Handbook of differential equations. Stationary partial differential equations [Volume 2] 9780080461076, 9780444520456, 0444520457

Citation preview

H ANDBOOK OF D IFFERENTIAL E QUATIONS S TATIONARY PARTIAL D IFFERENTIAL E QUATIONS VOLUME II

This page intentionally left blank

H ANDBOOK OF D IFFERENTIAL E QUATIONS S TATIONARY PARTIAL D IFFERENTIAL E QUATIONS Volume II

Edited by

M. CHIPOT Institute of Mathematics, University of Zürich, Zürich, Switzerland

P. QUITTNER Department of Applied Mathematics and Statistics, Comenius University, Bratislava, Slovak Republic

2005 ELSEVIER Amsterdam • Boston • Heidelberg • London • New York • Oxford Paris • San Diego • San Francisco • Singapore • Sydney • Tokyo

ELSEVIER B.V. Radarweg 29 P.O. Box 211 1000 AE Amsterdam The Netherlands

ELSEVIER Inc. 525 B Street, Suite 1900 San Diego, CA 92101-4495 USA

ELSEVIER Ltd The Boulevard, Langford Lane Kidlington, Oxford OX5 1GB UK

ELSEVIER Ltd 84 Theobalds Road London WC1X 8RR UK

© 2005 Elsevier B.V. All rights reserved. This work is protected under copyright by Elsevier B.V., and the following terms and conditions apply to its use: Photocopying Single photocopies of single chapters may be made for personal use as allowed by national copyright laws. Permission of the Publisher and payment of a fee is required for all other photocopying, including multiple or systematic copying, copying for advertising or promotional purposes, resale, and all forms of document delivery. Special rates are available for educational institutions that wish to make photocopies for non-profit educational classroom use. Permissions may be sought directly from Elsevier’s Rights Department in Oxford, UK; phone: (+44) 1865 843830, fax: (+44) 1865 853333, e-mail: [email protected]. Requests may also be completed on-line via the Elsevier homepage (http://www.elsevier.com/locate/permissions). In the USA, users may clear permissions and make payments through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA; phone: (+1) (978) 7508400, fax: (+1) (978) 7504744, and in the UK through the Copyright Licensing Agency Rapid Clearance Service (CLARCS), 90 Tottenham Court Road, London W1P 0LP, UK; phone: (+44) 20 7631 5555; fax: (+44) 20 7631 5500. Other countries may have a local reprographic rights agency for payments. Derivative Works Tables of contents may be reproduced for internal circulation, but permission of the Publisher is required for external resale or distribution of such material. Permission of the Publisher is required for all other derivative works, including compilations and translations. Electronic Storage or Usage Permission of the Publisher is required to store or use electronically any material contained in this work, including any chapter or part of a chapter. Except as outlined above, no part of this work may be reproduced, stored in a retrieval system or transmitted in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without prior written permission of the Publisher. Address permissions requests to: Elsevier’s Rights Department, at the fax and e-mail addresses noted above. Notice No responsibility is assumed by the Publisher for any injury and/or damage to persons or property as a matter of products liability, negligence or otherwise, or from any use or operation of any methods, products, instructions or ideas contained in the material herein. Because of rapid advances in the medical sciences, in particular, independent verification of diagnoses and drug dosages should be made.

First edition 2005 Library of Congress Cataloging in Publication Data A catalog record is available from the Library of Congress. British Library Cataloguing in Publication Data A catalogue record is available from the British Library.

ISBN: 0 444 52045 7 Set ISBN: 0 444 51743 x The paper used in this publication meets the requirements of ANSI/NISO Z39.48-1992 (Permanence of Paper). Printed in The Netherlands.

Preface This handbook is Volume II in a series devoted to stationary partial differential equations. Similarly as Volume I, it is a collection of self-contained, state-of-the-art surveys written by well-known experts in the field. The topics covered by this handbook include existence and multiplicity of solutions of superlinear elliptic equations, bifurcation phenomena, problems with nonlinear boundary conditions, nonconvex problems of the calculus of variations and Schrödinger operators with singular potentials. We hope that these surveys will be useful for both beginners and experts and speed up the progress of corresponding (rapidly developing and fascinating) areas of mathematics. We thank all the contributors for their clearly written and elegant articles. We also thank Arjen Sevenster and Andy Deelen at Elsevier for efficient collaboration. M. Chipot and P. Quittner

v

This page intentionally left blank

List of Contributors Bartsch, T., Universität Giessen, 35392 Giessen, Germany (Ch. 1) Dacorogna, B., EPFL, 1015 Lausanne, Switzerland (Ch. 2) Du, Y., University of New England, Armidale, NSW 2351, Australia (Ch. 3) López-Gómez, J., Universidad Complutense de Madrid, 28040 Madrid, Spain (Ch. 4) Melgaard, M., Uppsala University, S-751 06 Uppsala, Sweden (Ch. 6) Rossi, J.D., Universidad de Buenos Aires, 1428 Buenos Aires, Argentina (Ch. 5) Rozenblum, G., Chalmers University of Technology, and University of Gothenburg, S-412 96 Gothenburg, Sweden (Ch. 6) Solimini, S., Politecnico di Bari, 70125 Bari, Italy (Ch. 7) Wang, Z.-Q., Utah State University, Logan, UT 84322, USA (Ch. 1) Willem, M., Université catholique de Louvain, 1348 Louvain-la-Neuve, Belgium (Ch. 1)

vii

This page intentionally left blank

Contents Preface List of Contributors Contents of Volume I

v vii xi

1. The Dirichlet Problem for Superlinear Elliptic Equations T. Bartsch, Z.-Q. Wang and M. Willem 2. Nonconvex Problems of the Calculus of Variations and Differential Inclusions B. Dacorogna 3. Bifurcation and Related Topics in Elliptic Problems Y. Du 4. Metasolutions: Malthus versus Verhulst in Population Dynamics. A Dream of Volterra J. López-Gómez 5. Elliptic Problems with Nonlinear Boundary Conditions and the Sobolev Trace Theorem J.D. Rossi 6. Schrödinger Operators with Singular Potentials G. Rozenblum and M. Melgaard 7. Multiplicity Techniques for Problems without Compactness S. Solimini

1 57 127

211

311 407 519

Author Index

601

Subject Index

609

ix

This page intentionally left blank

Contents of Volume I Preface List of Contributors

v vii

1. Solutions of Quasilinear Second-Order Elliptic Boundary Value Problems via Degree Theory C. Bandle and W. Reichel 2. Stationary Navier–Stokes Problem in a Two-Dimensional Exterior Domain G.P. Galdi 3. Qualitative Properties of Solutions to Elliptic Problems W.-M. Ni 4. On Some Basic Aspects of the Relationship between the Calculus of Variations and Differential Equations P. Pedregal 5. On a Class of Singular Perturbation Problems I. Shafrir 6. Nonlinear Spectral Problems for Degenerate Elliptic Operators P. Takáˇc 7. Analytical Aspects of Liouville-Type Equations with Singular Sources G. Tarantello 8. Elliptic Equations Involving Measures L. Véron

1 71 157

235 297 385 491 593

Author Index

713

Subject Index

721

xi

This page intentionally left blank

CHAPTER 1

The Dirichlet Problem for Superlinear Elliptic Equations Thomas Bartsch Mathematisches Institut, Universität Giessen, Arndtstrasse 2, 35392 Giessen, Germany E-mail: [email protected]

Zhi-Qiang Wang Department of Mathematics and Statistics, Utah State University, Logan, UT 84322, USA E-mail: [email protected]

Michel Willem Institut de mathématique pure et appliquée, Université catholique de Louvain, 1348 Louvain-la-Neuve, Belgium E-mail: [email protected]

Contents 0. Introduction . . . . . . . . . . . . . . . . . . . . . . 0.1. Conditions on the nonlinearity . . . . . . . . . 1. Positive solutions . . . . . . . . . . . . . . . . . . . 1.1. Existence of positive solutions . . . . . . . . . 1.2. Uniqueness of positive solutions . . . . . . . . 1.3. The Nehari manifold . . . . . . . . . . . . . . 1.4. Existence of ground states . . . . . . . . . . . 1.5. Symmetry of the ground state solution . . . . . 1.6. Multiple positive solutions . . . . . . . . . . . 1.7. The method of moving planes . . . . . . . . . 1.8. A priori bounds for positive solutions . . . . . 2. Nodal solutions on bounded domains . . . . . . . . 2.1. A natural constraint . . . . . . . . . . . . . . . 2.2. Localizing critical points . . . . . . . . . . . . 2.3. Upper bounds on the number of nodal domains

. . . . . . . . . . . . . . .

HANDBOOK OF DIFFERENTIAL EQUATIONS Stationary Partial Differential Equations, volume 2 Edited by M. Chipot and P. Quittner © 2005 Elsevier B.V. All rights reserved 1

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

3 4 4 4 10 13 16 17 17 18 21 22 22 24 27

2

T. Bartsch et al.

2.4. The existence of nodal solutions . . . . . . . . . . . . . . . . . . . . . . 2.5. Geometric properties of least energy nodal solutions on radial domains 2.6. Multiple nodal solutions on a bounded domain . . . . . . . . . . . . . . 2.7. Perturbed symmetric functionals . . . . . . . . . . . . . . . . . . . . . . 3. Problems on the entire space . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1. The compact case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2. The radially symmetric case . . . . . . . . . . . . . . . . . . . . . . . . 3.3. The steep potential well case . . . . . . . . . . . . . . . . . . . . . . . . 3.4. Ground state solutions for bounded potentials . . . . . . . . . . . . . . . 3.5. More on periodic potentials . . . . . . . . . . . . . . . . . . . . . . . . . 3.6. Strongly indefinite potentials . . . . . . . . . . . . . . . . . . . . . . . . Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

28 31 33 35 35 36 39 40 43 45 51 51 52

The Dirichlet problem for superlinear elliptic equations

3

0. Introduction Boundary value problems for nonlinear elliptic partial differential equations have been a major focus of research in nonlinear analysis for decades. In this survey we discuss semilinear equations like


−
u + a(x)u = f (x, u), u=0

x ∈ Ω, on ∂Ω,

(0.1)

where Ω is a domain in RN , N
2, and f : Ω × R → R is superlinear, that is, f (x, t)/t → ∞ as |t| → ∞. The model nonlinearity is the homogeneous function f (x, t) = |t|p−2 t

with p > 2.

(0.2)

The continuation method or other classical methods based on the Leray–Schauder degree do not apply easily to (0.1) because there are no a priori bounds for the solutions. This is inherent to superlinear nonlinearities. In fact, for the model nonlinearity (0.2) with p > 2, p < 2N/(N − 2) if N
3, a ∈ L∞ (Ω), and bounded Ω, there exist infinitely many solutions which are unbounded in the H 1 norm. On the other hand, if Ω = RN is star-shaped, a(x) ≡ const
0, f is as in (0.2) with p
2N/(N − 2), N
3, then (0.1) has no solution except the trivial one u ≡ 0. Here we know a posteriori that the solutions are bounded but the Leray–Schauder methods do not apply due to a lack of compactness. After some initial work during the 1960s and early 1970s, Ambrosetti and Rabinowitz established in the seminal paper [5] several variational methods to obtain solutions of (0.1) on bounded domains, most notably the mountain pass theorem and variations thereof. These methods have been refined and extended to deal, for instance, with unbounded domains or the critical exponent case f (x, t) = b(x)|t|4/(N −2) t which is closely related to the Yamabe problem from differential geometry. With these methods more complicated partial differential equations with variational structure can now be investigated. Moreover, qualitative properties of the solutions have been discovered in recent years, in particular, on the nodal structure and the symmetry of the solutions. The goal of this chapter is to present some basic ideas in a simple setting and to survey selected results on the Dirichlet problem (0.1) with superlinear nonlinearity. The chapter consists of three sections. In Section 1 we deal with positive solutions of (0.1), in Section 2 with sign-changing solutions on bounded domains and in Section 3 we treat the unbounded domain Ω = RN . No effort is being made to be as general as possible. Neither did we try to write a comprehensive survey on (0.1). For example, we do not present results on the bifurcation of solutions nor for the p-Laplace operator, nor do we treat singularly perturbed equations in detail. These topics require separate surveys. Fortunately, there are a number of well written monographs about (0.1) where the reader can find additional information and further references. We would like to mention the books by Rabinowitz [79], Struwe [87], Chang [41], Ghoussoub [52], Kavian [56], Schechter [81], Willem [95], Chabrowski [39,40], Kielhöfer [57]. We concentrate on topics not being covered in these monographs, although a certain overlap cannot be avoided for natural reasons. Of course, the choice of topics is also influenced by our own research interests.

4

T. Bartsch et al.

0.1. Conditions on the nonlinearity For the convenience of the reader we list here conditions on f which we use at various places in the chapter. The critical exponent is defined by 2∗ =




+∞ 2N N −2

if N = 2, if N
3.

t Given f : Ω × R → R we let F : Ω × R → R, F (x, t) := 0 f (x, s) ds, be the primitive of f . (f0 ) f : Ω × R → R is a Carathéodory function with f (x, t) = o(t) as t → 0. There exist C > 0 and p  2∗ such that |f (x, t)|  C(|t| + |t|p−1 ) for all x ∈ Ω, t ∈ R. ′ (f0 ) f : Ω × R → R is differentiable in t ∈ R, and the derivative ft is a Carathéodory function with ft (x, 0) = 0. There exist C > 0 and p  2∗ such that |ft (x, t)|  C(1 + |t|p−2 ) for all x ∈ Ω, t ∈ R. (f1 ) For every x ∈ Ω the function R \ {0} → R, t → f (x, t)/|t|, is strictly increasing. (f1′ ) ft (x, t) > f (x, t)/t for every x ∈ Ω and every t = 0. (f2 ) There exist R
0 and θ > 2 such that 0 < θF (x, t)  f (x, t)t for all x ∈ Ω, |t| > R. If Ω = RN it is required that R = 0. (f3 ) lim|t|→∞ F (x, t)/t 2 = +∞, uniformly in x. (f4 ) There exists m > 0 so that t → f (x, t) + mt is strictly increasing for all x ∈ Ω. (f5 ) inft=0 f (x, t)/t > −∞. 1. Positive solutions 1.1. Existence of positive solutions We consider first the problem  p−1 , x ∈ Ω,   −
u + a(x)u = u u > 0, x ∈ Ω,   u=0 on ∂Ω,

(1.1)

where Ω is a smooth domain in RN , a ∈ L∞ (Ω) and 2 < p  2∗ . In order to obtain a solution of (1.1) we assume that −
+ a is positive, i.e., there exists c > 0 such that, for every u ∈ H01 (Ω), 

Ω

|∇u|2 + a(x)u2 dx
c



u2 dx.

(1.2)

Ω

Our main tool is the Rellich compactness theorem. T HEOREM 1.1. Let Ω be bounded and let 1  p < 2∗ . Then the injection H01 (Ω) ⊂ Lp (Ω) is compact.

The Dirichlet problem for superlinear elliptic equations

5

T HEOREM 1.2. Assume that Ω is bounded, 2 < p < 2∗ and (1.2) is satisfied. Then there is a solution of (1.1). P ROOF. By Theorem 1.1 and lower semicontinuity, it is easy to verify that µ=



inf

v∈H01 (Ω) vLp =1

Ω

|∇v|2 + a(x)v 2 dx

(1.3)

is achieved by some v. ¯ After replacing v¯ by |v|, ¯ we may assume that v¯
0. It follows from the Lagrange multiplier rule that −
v¯ + a(x)v¯ = µv¯ p−1 . A solution of (1.1) is then given by u¯ = µ1/(p−2) v. ¯ Indeed, u¯ > 0 on Ω by the strong maximum principle.  R EMARK 1.3. There are other ways to prove Theorem 1.2. Instead of minimizing as in (1.3) one can minimize the functional Φ(u) =

1 2



Ω

|∇u|2 + a(x)u2 dx −

1 p



Ω

|u|p dx

on the Nehari manifold  N = u ∈ H01 (Ω): u = 0, Φ ′ (u)u = 0


  = u ∈ H01 (Ω): u = 0, |∇u|2 + a(x)u2 dx = |u|p dx . Ω

Ω

A critical point of the constrained functional Φ|N is a critical point of Φ, hence a solution of (1.1). The map,  v ∈ H01 (Ω): vLp = 1 → N ,

v →



u →

u . uLp

Ω

2

2

|∇v| + a(x)v dx

1/(p−2)

v,

is a diffeomorphism with inverse  N → v ∈ H01 (Ω): vLp = 1 ,

It maps the minimizer v¯ of (1.3) to the minimizer u¯ of Φ on N . The solution can also be obtained via the mountain pass theorem from [5]. In fact,

 Φ(u) ¯ = inf max Φ(tu) = inf max Φ γ (t) , u=0 t
0

γ ∈Γ t∈[0,1]

6

T. Bartsch et al.

where Γ consists of all continuous paths γ : [0, 1] → H01 (Ω) with γ (0) = 0 and Φ(γ (1)) < 0. These three different approaches are equally valid for (1.1). They allow different generalizations. The mountain pass approach leads to the most general existence results for positive solutions of −
u + a(x)u = f (x, u) with Dirichlet boundary conditions. This approach is most widely used in the literature. Minimizing over the Nehari manifold requires more conditions on the nonlinearity f . When these are satisfied one can find nodal solutions on the Nehari manifold and obtain useful additional information, in particular on the nodal structure and the symmetry. Minimizing as in (1.3) only makes sense for homogeneous nonlinearities. The critical case p = 2∗ and the supercritical case p > 2∗ are more delicate. T HEOREM 1.4. Suppose that N
3, p
2∗ , Ω = RN is star-shaped with smooth boundary, and a(x) ≡ λ
0. Then there is no solution of (1.1). P ROOF. By the Pohozaev identity [78], if
−
u = f (u) in Ω, u=0 on ∂Ω, then 

  1 N −2 2 |∇u| dx = |∇u|2 σ · ν dσ, F (u) dx − N 2 2 ∂Ω Ω Ω u where F (u) = 0 f (s) ds and ν denotes the unit outward normal to ∂Ω. For a solution u of (1.1) we therefore obtain −λ



Ω

u2 dx +



N −2 N − p 2



Ω

|u|p dx =



∂Ω

|∇u|2 σ · ν dσ
0. 2

It follows that u = 0 or λ  0, hence λ = 0. If λ = 0, then p = 2∗ and ∇u = 0 on ∂Ω, so by (1.1)   0 = −
u dx = up−1 dx, Ω

and therefore, u = 0.

Ω



R EMARK 1.5. In the situation of Theorem 1.4, there is, in fact, no positive nor signchanging solution u = 0 of
−
u + λu = |u|p−2 u, x ∈ Ω, (1.4) u=0 on ∂Ω.

As in the proof of Theorem 1.4, for a solution u = 0 to exist we must have λ = 0, p = 2∗ , and then ∇u = 0 on ∂Ω. Now u = 0 follows from the unique continuation principle.

The Dirichlet problem for superlinear elliptic equations

7

We assume now that Ω is bounded and a(x) ≡ −λ, where 0 < λ < λ1 (Ω) and λ1 (Ω) = min

v∈H01 (Ω) v 2 =1 L



Ω

|∇v|2 dx.

We consider the minimization problem Sλ =

inf

v∈H01 (Ω) v 2∗ =1 L



Ω

|∇v|2 − λv 2 dx.

(1.5)

In order to solve this problem, we need two tools: the following Brezis–Lieb lemma from [31] and the strict inequality Sλ < S, where S=

inf

v∈H 1 (RN ) v 2∗ =1 L



RN

|∇v|2 dx.

L EMMA 1.6. Let (un ) be a bounded sequence in Lp (Ω), 1  p < ∞, such that un → u a.e. on Ω. Then

p p  p lim un Lp − un − uLp = uLp .

n→∞

L EMMA 1.7. Let N
4 and λ > 0. Then Sλ < S. P ROOF. The instanton U (x) =

[N (N − 2)](N −2)/4 [1 + |x|2 ](N −2)/2

(1.6)

is a minimizer for S. Since U |Ω ∈ / H01 (Ω), we have to use a truncation ψ. We can assume that Bρ (0) ⊂ Ω. Let ψ ∈ D(Ω), ψ
0, be such that ψ ≡ 1 on Bρ (0). Using Uε (x) = ψ(x)ε

(2−N )/2

 x U ε

an asymptotic analysis shows that Sλ < S.



R EMARK 1.8. When N = 3, the situation is more delicate. Consider the unit ball Ω = B1 (0) ⊂ R3 . Then we have 00 on RN ,   u(x) → 0, |x| → ∞. The corresponding minimization problem is  Sp = inf |∇v|2 + v 2 dx. v∈H 1 (RN ) vLp =1

RN

Let us recall that H 1 (RN ) = H01 (RN ). We shall use the Schwarz symmetrization and radial Sobolev inequalities. Let VN = m(B1 (0)) be the Lebesgue measure of the unit ball in RN . The Schwarz symmetrization of a measurable subset A of RN is defined by A∗ = BR (0),

 where R is given by m BR (0) = VN R N = m(A).

The Schwarz symmetrization of a measurable function u : RN → [0, ∞[ is defined by  u∗ (x) = sup t > 0: x ∈ {u > t}∗ .

T HEOREM 1.11. Let u : RN → [0, ∞[ be measurable and let f : [0, ∞[ → R be continuous. Then  

 f u∗ dx = f (u) dx. RN

RN

Let 1  p < ∞ and u ∈ W 1,p (RN ), u
0. Then the Pólya–Szegö inequality holds:  ∗ ∇u 

Lp

 ∇uLp .

See [34] or [96] for a simple proof. We denote by Hr1 (RN ) the space of radial functions of H 1 (RN ). Let us recall that a function u is radial if u = u(|x|). The Schwarz symmetrization of a measurable function is radial. The following results are due to Strauss [85]. L EMMA 1.12. Let N
2. There exists c(N ) > 0 such that, for every u ∈ Hr1 (RN ),   1/2 u(x)  c(N )u1/2 ∇uL2 |x|(1−N )/2 L2

a.e. on RN .

T HEOREM 1.13. Let 2 < p < 2∗ . Then there is a radial solution of problem (1.7).

10

T. Bartsch et al.

P ROOF. Let (vn ) ⊂ H 1 (RN ) be a minimizing sequence for Sp : vn Lp = 1, 

RN

|∇vn |2 + vn2 dx → Sp .

By Theorem 1.11, we can replace (vn ) by (vn∗ ). By Theorem 1.1 and Lemma 1.12, we can assume, going if necessary to a subsequence:

 in H 1 RN ,

 vn∗ → v in Lp RN ,

vn∗ ⇀ v

where v is a radial function. Clearly, v is a minimizer for Sp and (1.7) is solvable.



The existence of nonradial entire solutions u ∈ H 1 (RN ) of (1.1) and more general equations will be discussed in Section 3.

1.2. Uniqueness of positive solutions The problem of uniqueness of positive solutions is mostly solved for symmetric domains and is closely related to the symmetry of solutions. Let us first recall a celebrated result proved in 1979 by Gidas, Ni and Nirenberg [53].  T HEOREM 1.14. Let Ω be the unit ball in RN . Assume that f is C 1 and u ∈ C 2 (Ω) satisfies   −
u = f (u), u > 0,  u=0

x ∈ Ω, x ∈ Ω, on ∂Ω.

Then u is a radial function and u′ (r) is negative. Consider the problem  p−1 , x ∈ Ω = B (0),  1  −
u + λu = u u > 0, x ∈ Ω,   u=0 on ∂Ω,

where Ω = B1 (0) is the unit ball in RN . Uniqueness is proved when λ = 0, 2 < p < 2∗ ,

Gidas, Ni and Nirenberg [53], 1979,

λ > 0, 2 < p < 2 ,

Kwong [59], 1989,

λ < 0, 2 < p  2∗ ,

Srikanth [84], 1993.

∗

(1.8)

The Dirichlet problem for superlinear elliptic equations

11

Concerning problem (1.7), uniqueness for 2 < p < 2∗ is proved in [59] after the pioneering work of Coffman [42] in 1972. The situation differs when Ω is an annulus, Ω = {x ∈ RN : r < x < R} for some R > r > 0. Let us recall a particular case of the principle of symmetric criticality proved in 1979 by Palais. D EFINITION 1.15. The action of a topological group G on a normed space X is a continuous map G × X → X : (g, u) → gu such that 1 · u = u, (gh)u = g(hu), u → gu is linear. The action is isometric if gu ≡ u. The orbit of an element u ∈ X is the set Gu := {gu: g ∈ G}, and the space of fixed points is defined by  Fix(G) = {u ∈ X: gu = u for all g ∈ G} = u ∈ X: Gu = {u} . A function ϕ : X → R is invariant if ϕ ◦ g = ϕ for every g ∈ G.

E XAMPLE 1.16. Assume that Ω is invariant by rotations: for every g ∈ SO(N ), gΩ = Ω. The action of SO(N ) on H01 (Ω) is defined by

 gu(x) = u g −1 x .

1 (Ω) of radial functions in H 1 (Ω). From TheoThe space Fix(SO(N )) is the space H0,r 0 1 (Ω) ⊂ Lp (Ω) is compact for rem 1.1 and Lemma 1.12, it follows that the injection H0,r 2 < p < 2∗ . Moreover, if Ω is an annulus  Ω = x ∈ RN : ρ < |x| < R

or an exterior domain  Ω = x ∈ RN : ρ < |x| ,

1 ⊂ Lp (Ω) is compact for 2 < p  ∞. the injection H0,r

Let us recall that, for every open subset Ω of RN , S(Ω) :=

inf

u∈H01 (Ω) u2∗ =1

∇u22 = S

12

T. Bartsch et al.

and S(Ω) is never achieved except when Ω = RN . Since the “radial infimum” is achieved when Ω is an annulus, we have S=

inf

u∈H01 (Ω) u 2∗ =1 L

∇u2L2 < µ2∗ ,r =

min ∇u2L2 .

1 (Ω) u∈H0,r u 2∗ =1 L

We deduce µp = min ∇u2L2 < µp,r = u∈H01 (Ω) uLp =1

min ∇u2L2

1 (Ω) u∈H0,r uLp =1

for 2∗ − ε < p < 2∗ .

(1.9)

Using the symmetric criticality principle, we construct, for 2 < p < ∞, a radial solution of the problem  p−1 ,  −
u = u  u > 0, u=0

 x ∈ Ω = x ∈ RN : r < x < R ,

x ∈ Ω, on ∂Ω.

(1.10)

For 2∗ − ε < p < 2∗ inequality (1.9) yields a nonradial solution of problem (1.10). Thus we have proved the following theorem. T HEOREM 1.17. For 2∗ − ε < p < 2∗ , problem (1.10) has at least two solutions, one radial and one nonradial. Moreover, the least energy solution is nonradial. The above result is due to Brezis and Nirenberg [32]. Following their work, there are related results for multiple positive nonradial solutions of semilinear elliptic equations on expanding annular domains, by Coffman [43] for N = 2, by Li [62] for N
4, and by Byeon [35] (and Catrina and Wang [38] independently) for N = 3. We set Ωa = {x ∈ RN : a < |x| < a + 1} and consider  p−1 ,   −
u + u = u u > 0,   u = 0,

x ∈ Ωa , x ∈ Ωa , x ∈ ∂Ωa ,

(1.11)

where N
2, 2 < p < 2∗ .

T HEOREM 1.18. The number of not rotationally equivalent, nonradial solutions of (1.11) tends to infinity as a → ∞. In [38] the exact symmetry of these nonradial positive solutions was examined and positive solutions having a prescribed symmetry can be constructed from a local minimization procedure. The article [93] has results for the critical exponent problems. More results on multiple positive solutions will be mentioned in Section 1.6.

The Dirichlet problem for superlinear elliptic equations

13

1.3. The Nehari manifold Let Ω ⊂ RN be a (not necessarily bounded) smooth domain. We consider the problem


−
u + a(x)u = f (x, u),

u=0

x ∈ Ω,

(1.12)

on ∂Ω,

where a ∈ L∞ (Ω). Concerning the nonlinearity we assume: (f0 ) f : Ω × R → R is a Carathéodory function with f (x, t) = o(t) as t → 0, uniformly in x. There exist C > 0 and p  2∗ such that |f (x, t)|  C(|t| + |t|p−1 ) for all x ∈ Ω, t ∈ R. (f1 ) For every x ∈ Ω the function R \ {0} → R, t → f (x, t)/|t|, is strictly increasing. By (f0 ) the functional Φ :E

1 Φ(u) = 2

= H01 (Ω) → R,



Ω

2

2

|∇u| + a(x)u dx −



F (x, u) dx,

Ω

t where F (x, t) = 0 f (x, s) ds, is well defined and C 1 . Critical points of Φ are weak solutions of (1.12). The Nehari manifold is defined by  N = u ∈ E \ {0}: Φ ′ (u)u = 0 .

We define also

 N± = {u ∈ N : ±u
0} and β± := inf Φ(u): u ∈ N± .

It is clear that N contains all nontrivial critical points of Φ. Moreover, in order to find the least energy positive (resp. negative) solution of (1.12), it suffices to minimize Φ on N+ (resp. N− ). We begin with a geometric description of N . We assume that f satisfies (f0 ) and (f1 ), or the following differentiable versions: (f0′ ) f : Ω × R → R is differentiable in t ∈ R with f (x, t) = o(t) as t → 0, uniformly in x. The derivative ft is a Carathéodory function. There exist C > 0 and p  2∗ such that |ft (x, t)|  C(1 + |t|p−2 ) for all x ∈ Ω, t ∈ R. ′ (f1 ) ft (x, t) > f (x, t)/t for every x ∈ Ω and every t = 0. Clearly, (f0′ ) and (f1′ ) imply (f0 ) and (f1 ). For u ∈ SE := {u ∈ E: u = 1}, the map, ψu : R+ → R, 1 λ → Φ ′ (λu)u = λ



Ω

|∇u|2 + a(x)u2 dx −

is strictly decreasing by (f1 ). The set  U := u ∈ SE: ψu (λ) < 0 for some λ > 0



Ω

f (x, λu) u dx, λ

(1.13)

14

T. Bartsch et al.

is an open subset of SE. For u ∈ U there exists a unique λu > 0 such that ψu (λu u) = 0, that is, λu u ∈ N . P ROPOSITION 1.19. (a) If f satisfies (f0 ) and (f1 ) then the map h:U → N,

u → λu u

is a homeomorphism with inverse N → U , v → v/v. (b) If f satisfies (f0′ ) and (f1′ ) then N is a C 1 -manifold with tangent space Tu N = {v ∈ E: Φ ′′ (u)[u, v] + Φ ′ (u)v = 0}. The map h is a C 1 -diffeomorphism. If u ∈ N is a critical point of the constrained functional Φ|N , then u is a critical point of Φ. P ROOF. (a) is a simple consequence of the fact that ψu is strictly decreasing and that ψu (λ) is continuous in (u, λ). (b) Here Φ ∈ C 2 (E) and the implicit function theorem applied to the map U × R+ ,

(u, λ) → ψu (λ)

yields that u → λu is C 1 . The claims follow because Tu N is transversal to R+ u.



We can describe the set U if f is superlinear: (f2 ) There exist R > 0 and θ > 2 such that 0 < θF (x, t)  f (x, t)t for all x ∈ Ω, |t| > R. L EMMA 1.20. Suppose (f0 ), (f1 ) and (f2 ) hold. Then U = {u ∈ SE: Q(u) > 0}, where Q(u) := Ω (|∇u|2 + a(x)u2 ) dx. If −
+ a is positive then U = SE and N is bounded away from 0. The easy proof is left to the reader. T HEOREM 1.21. If (f0 ) and (f1 ) hold then β± = infu∈N Φ(u)
0 and every minimizer of Φ on N+ (resp. N− ) is a positive (resp. negative) solution of (1.12). We do not claim that β± is achieved. This requires additional conditions on f and a; see Theorem 1.23. Since N is not, in general, a differentiable manifold, the Lagrange multiplier rule is not applicable. We shall use a general deformation lemma. By definition Φ c = Φ −1 (]−∞, c]) and Sδ = {u ∈ X: dist(u, S)  δ}. L EMMA 1.22. Let X be a Banach space, Φ ∈ C 1 (X, R), S ⊂ X, c ∈ R, ε > 0, δ > 0, such that

 ∀u ∈ Φ −1 [c − 2ε, c + 2ε] ∩ S2δ :

Then there exists η ∈ C([0, 1] × X, X) such that

 ′  8ε Φ (u)
. δ

The Dirichlet problem for superlinear elliptic equations

(i) (ii) (iii) (iv) (v) (vi)

15

if t = 0 or if u ∈ / Φ −1 ([c − 2ε, c + 2ε]) ∩ S2δ then η(t, u) = u; c+ε η(1, Φ ∩ S) ⊂ Φ c−ε ; η(t, ·) is a homeomorphism of X ∀t ∈ [0, 1]; η(t, u) − u  δ ∀u ∈ X, ∀t ∈ [0, 1]; Φ(η(·, u)) is nonincreasing ∀u ∈ X; Φ(η(t, u)) < c ∀u ∈ Φ c ∩ Sδ , ∀t ∈ [0, 1].

A proof can be found in [95], Lemma 1.4. P ROOF OF T HEOREM 1.21. The inequality β±
0 follows from the fact that the map ψu from (1.13) is strictly decreasing. Now let u ∈ N + be such that Φ(u) = β+ . We shall prove that Φ ′ (u) = 0. It follows from assumption (f1 ) that Φ(su) < Φ(u) = β+

for 0 < s = 1.

If Φ ′ (u) = 0, then there exists δ > 0 and λ > 0 such that v − u  3δ

⇒

Φ ′ (v)
λ.

Clearly β0 = max{Φ(u/2), Φ(3u/2)} < β+ . For ε = min{(β+ − β0 )/2, (λδ)/8} and S = Bδ (u), Lemma 1.22 yields a deformation η such that • η(1, v) = v if v ∈ / Φ −1 ([β+ − 2ε, β+ + 2ε]), β +ε + • η(1, Φ ∩ B(u, δ)) ⊂ Φ β+ −ε , • Φ(η(1, v))  Φ(v) for all v ∈ H01 (Ω). Let us define, for 1/2 < s < 3/2,  h(s) = max η(1, su), 0 ∈ H01 (Ω).

It is clear that

max

1/2s3/2

 Φ h(s) < β+ .

We shall prove that h([1/2, 3/2]) ∩ N + = ∅, contradicting the definition of β+ . Since    1 1 h = Φ h 2 2    3 3 ′ Φ h h = 2 2 ′

 1 ′ u u > 0, Φ 2 2  3 ′ 3u u < 0, Φ 2 2

the existence of s ∈ (1/2, 3/2) with Φ ′ (h(s))h(s) = 0, i.e., h(s) ∈ N + , follows from the intermediate value theorem. 

16

T. Bartsch et al.

1.4. Existence of ground states In order to prove the existence of a minimizer of Φ on N+ , we assume (f3 ) lim|t|→∞ F (x, t)/t 2 = +∞, uniformly in x. T HEOREM 1.23. Suppose that Ω is bounded and that f satisfies (f0 ) with p < 2∗ , (f1 ) and (f3 ). Moreover, suppose that −
+ a is positive. Then there exists a minimizer of Φ on N+ and, hence, a positive solution of (1.12). P ROOF. Let (un ) ⊂ N+ be a minimizing sequence: Φ(un ) → β+ . Let us define tn = un  and vn = un /tn . We can assume that vn ⇀ v in H01 (Ω). Since, for every R > 0,    R2 2 1+ avn dx − F (x, Rvn ) dx = Φ(Rvn )  Φ(un ), 2 Ω Ω we have    R2 2 1+ a(x)v dx − F (x, Rv) dx  β+ , 2 Ω Ω so that v = 0. If (tn ) is unbounded, we can assume that tn → +∞. We obtain, from (f3 ) and Fatou’s lemma, the contradiction  a(x)vn2 F (x, tn vn ) + dx lim inf − 2 tn2 Ω   1 a(x)vn2 F (x, tn vn )  lim + dx = . − 2 n→∞ Ω 2 2 tn

+∞ =



It follows that (un ) is bounded in H01 (Ω). Going if necessary to a subsequence, we can assume that un ⇀ u in H01 (Ω). By (f0 ), since (un ) ⊂ N+ , 

Ω

f (u)u dx = lim



n→∞ Ω

= lim

n→∞



f (un )un dx

Ω

a(x)u2n dx + |∇un |2 > 0.

In particular, u = 0. Since u
0, there exists t
0 such that tu ∈ N + and Φ(tu)  lim inf Φ(tun )  lim inf Φ(un ) = β+ . n→∞

n→∞

Hence tu is a minimizer of Φ on N+ .



Theorem 1.23 is due to Liu and Wang [67]. Note that (f3 ) is weaker than (f2 ). Efforts in weakening (f2 ) have been made in [46,82] (see also the references therein).

The Dirichlet problem for superlinear elliptic equations

17

1.5. Symmetry of the ground state solution In this section we assume that Ω is invariant by rotations: gΩ = Ω for every g ∈ SO(N ). Theorem 1.17 shows that, even when f is independent of x, the ground state is, in general, not radial. However when f = f (|x|, u), a partial symmetry is always preserved by the ground state. We arbitrarily choose a fixed direction P in RN which we will refer to as the north pole direction. Let R > 0 and dσ denote the standard measure on ∂BR (0). The symmetrization A∗ of a measurable set A ⊂ ∂BR (0) is defined as the closed geodesic ball in ∂BR (0) centered at the north pole and whose dσ -measure equals that of A. The foliated Schwarz symmetrization B ∗ of a Borel set B ⊂ RN is defined on any sphere ∂BR (0) by

∗ B ∗ ∩ ∂BR (0) = B ∩ ∂BR (0)

and the foliated Schwarz symmetrization of a Borel function u : RN → [0, ∞[ is defined by 

∗ u∗ (x) = sup t > 0: x ∈ ∂BR (0) ∩ {u > t} .

See [83] for the Pólya–Szegö inequality corresponding to the foliated Schwarz symmetrization. The following result is proved in [19] by combining an elementary symmetrization, the polarization (see [34]) and the maximum principle. T HEOREM 1.24. We assume (f0 ), (f1 ) and (A1 ) f : Ω × R → R is Hölder continuous on Ω × [−R, R] for every R > 0, (A2 ) Ω is radially symmetric, (A3 ) a and f (·, t) are radial functions for every t ∈ R. Then every minimizer u of Φ on N+ or N− is a foliated Schwarz symmetric solution of (1.12). A related result can be found in [76], Theorem 3.1.

1.6. Multiple positive solutions Using a more topological argument Benci, Cerami and Passaseo [21,23] were able to establish the following result about the impact of the domain topology on the solution structure. Consider  p−1 , x ∈ Ω,   −
u + λu = u (1.14) u > 0, x ∈ Ω,   u = 0, x ∈ ∂Ω, on a bounded domain Ω ⊂ RN .

18

T. Bartsch et al.

T HEOREM 1.25. Suppose the Lusternik–Schnirelmann category of the domain satisfies cat(Ω)
2. (a) If 2 < p < 2∗ then for λ sufficiently large (1.14) has at least cat(Ω) + 1 solutions. (b) If λ
0 then for p < 2∗ sufficiently close to 2∗ (1.14) has at least cat(Ω) + 1 solutions. The case p = 2∗ is of special interest. It is closely related to the Yamabe problem from differential geometry. This critical case is analytically more difficult because the ∗ embedding H01 (Ω) ֒→ L2 (Ω) is not compact. As discussed in Section 1.1, Brezis and Nirenberg [32] obtained one solution for 0 < λ < λ1 if N
4, and, in the case N = 3, showed that there exists λ∗ ∈ [0, λ1 ) so that (1.14) has a solution for λ∗ < λ < λ1 . In the case λ = 0 and p = 2∗ , Bahri and Coron [8] obtained one solution if the domain has nontrivial homology: Hk (Ω; Z2 ) = 0 for some k
1. The existence of multiple positive solutions of (1.14) for p = 2∗ is not known. R EMARK 1.26. More results about the effect of the topology and geometry of the domains on the solutions structure have been given for the singularly perturbed nonlinear elliptic equation  2 p−1 ,   −ε
u + a(x)u = u u > 0,   u = 0,

x ∈ Ω, x ∈ Ω, x ∈ ∂Ω,

(1.15)

and variations thereof. In fact, Theorem 1.25 can easily be reformulated for (1.15). Typical results are concerned with the existence of multiple positive solutions and their limiting shape as ε → 0. This was first done for the least energy solutions in [75]. Singularly perturbed equations like (1.15) have been a very active area of research during the last fifteen years and the number of papers abound. A discussion of this topic goes beyond the scope of our survey.

1.7. The method of moving planes This section is related to Section 1.5. We consider the problem 

−
u = f (u) u>0 u=0

in Ω, in Ω, on ∂Ω,

(1.16)

where the domain Ω has some symmetry. We shall prove, by the method of moving planes, that, under some assumptions, all the solutions of (1.16) inherit the symmetry of Ω. The method of moving planes was introduced by Alexandrov in 1962 in the context of minimal surfaces and was used by Serrin in 1971 and Gidas, Ni and Nirenberg [53] in the study of semilinear elliptic equations. The method was extended and simplified by Berestycki and Nirenberg [26]. We describe a result of [26], following [29].

The Dirichlet problem for superlinear elliptic equations

19

We shall need the maximum principle for small domains. Let w be a solution of


−
w + c(x)w
0 in Ω, w
0 on ∂Ω.

(1.17)

The standard form of the maximum principle asserts that, if c(x)
0 in Ω then w(x)
0 in Ω. In Stampacchia’s form we use a weaker assumption. Let us recall that S=

inf

u∈H01 (Ω) u2∗ =1

∇u22

is independent of Ω. L EMMA 1.27. Assume that w ∈ H 1 (Ω) satisfies (1.17) with c− N/2 < S.

(1.18)

Then w
0 in Ω. P ROOF. It suffices to multiply (1.17) by w− , to integrate by parts and to use (1.18).



Assumption (1.18) is always satisfied if c− ∞ < ∞ and |Ω| is sufficiently small. T HEOREM 1.28. Let f be locally Lipschitz and let Ω be bounded, convex in some direction, say x1 , and symmetric with respect to the plane x1 = 0. Then any solution  of (1.16) is symmetric with respect to x1 and ∂u/∂x1 < 0 for x1 > 0 u ∈ C 2 (Ω) ∩ C(Ω) in Ω. P ROOF. Let us define  a = sup x1 : (x1 , y) ∈ Ω ,  Ωλ = (x1 , y) ∈ Ω: x1 > λ ,

0 < λ < a,

wλ (x) = u(2λ − x1 , y) − u(x1 , y),

x ∈ Ωλ .

The function wλ is well defined on Ωλ since Ω is convex in the direction x1 and symmetric with respect to the plane x1 = 0. We shall prove that wλ (x)
0 for 0 < λ < a, x ∈ Ωλ . In order to see this, we define cλ (x) =



f (u(x1 ,y))−f (u(2λ−x1 ,y)) wλ (x1 ,y)

0

if wλ (x) = 0, if wλ (x) = 0.

20

T. Bartsch et al.

The function wλ satisfies


−
wλ + cλ wλ = 0 wλ
0

in Ωλ , on ∂Ωλ .

(1.19)

Moreover, since u is bounded and f is locally Lipschitz,   sup sup cλ (x) < ∞.

(1.20)

0 0 using Hopf’s lemma.  R EMARK 1.29. (a) Theorem 1.14 follows directly from Theorem 1.28. It is interesting to note that Theorem 1.28 is applicable to domains like cubes. (b) The method of moving planes is very flexible and has been adapted to a large variety of problems. It is not possible to give a bibliography within this survey. The surveys by Berestycki [24] and by Brezis [29] contain many references. (c) With respect to Section 1.5 the assumptions on Ω and on f are somewhat stronger, but the results are applicable to any positive solution.

The Dirichlet problem for superlinear elliptic equations

21

1.8. A priori bounds for positive solutions Topological methods require the existence of a priori bounds for the set of all positive solutions. In this section we briefly describe particular cases of three classical results. We denote by λ1 the first eigenvalue of −
on H01 (Ω) and by e1 > 0 the corresponding eigenfunction. Throughout this section we assume that Ω is a smooth bounded domain in RN  × R+ → R+ is continuous. and that f : Ω The first result is due to Brezis and Turner [33]. The proof uses Hardy’s inequality. T HEOREM 1.30. Assume that lim inf u→∞

f (x, u) > λ1 u

 uniformly for x ∈ Ω,

and f (x, u) N +1  where α = . = 0 uniformly for x ∈ Ω, α u→∞ u N −1 lim

Then there exists c > 0 such that if u ∈ H01 (Ω) satisfies   −
u = f (x, u) + te1 u>0  u=0

in Ω, in Ω, on ∂Ω,

for some t
0, we have u∞  c. The second result is due to de Figueiredo, Lions and Nussbaum [51]. The proof uses the Pohozaev identity. T HEOREM 1.31. Assume that Ω is convex, f : R+ → R+ is locally Lipschitz and lim sup u→∞

f (u) > λ1 u

and

f (x, u) = 0, u→∞ uα lim

where α =

N . N −2

Then there exists c > 0 such that, if u satisfies   −
u = f (u) u>0  u=0

in Ω, in Ω, on ∂Ω,

we have uL∞  c.

The third result is due to Gidas and Spruck [54]. The proof uses a blow-up argument and Liouville theorems for the space RN or the half-space RN +.

22

T. Bartsch et al.

 → (0, ∞) such that T HEOREM 1.32. Assume that there exists a continuous function h : Ω lim

u→∞

f (x, u) = h(x) uα

 where 1 < α < uniformly for x ∈ Ω,

N +2 . N −2

 satisfies Then there exists c > 0 such that, if u ∈ C 2 (Ω) ∩ C(Ω)   −
u = f (x, u) in Ω, u>0 in Ω,  u=0 on ∂Ω,

we have u∞  c.

2. Nodal solutions on bounded domains In this section we report on recent results concerning nodal solutions of


−
u + a(x)u = f (x, u), u=0

x ∈ Ω, on ∂Ω.

(2.1)

Here Ω is a bounded smooth domain in RN , N
2, though some general results apply to unbounded domains. In the first two subsections we present two ideas how to localize critical points of the associated functional outside of the set of positive or negative functions. In the third subsection we give a nonlinear version of Courant’s nodal domain theorem for eigenfunctions of the Laplace operator. This gives an upper bound on the number of nodal domains of a solutions of (2.1) related to the min–max description of the critical value. In the Sections 2.4 and 2.5 we prove the existence and some properties of least energy nodal solutions. Finally, in Section 2.6 we study the existence of multiple nodal solutions.

2.1. A natural constraint In this subsection Ω may be unbounded. We consider the problem (2.1) with f satisfying (f0 ) and (f1 ) from Section 1.3. Recall the functional Φ : E = H01 (Ω) → R and the Nehari manifold  N = u ∈ E \ {0}: Φ ′ (u)u = 0

from Section 1.3. In Theorem 1.21 we showed that a minimizer of Φ on N + = {u ∈ N : u
0} is a positive solution of (2.1). In order to obtain nodal solutions, we consider the nodal Nehari set  S = u ∈ E: u+ ∈ N , u− ∈ N  = u ∈ E: u+ =  0 = u− , Φ ′ (u)u+ = 0 = Φ ′ (u)u− .

The Dirichlet problem for superlinear elliptic equations

23

Clearly, S ⊂ N contains the set of all nodal solutions. T HEOREM 2.1. Suppose (f0 ) and (f1 ) hold. Then β := inf Φ(S)
0, and every minimizer of Φ on S is a nodal solution of (2.1). In Section 2.4 we shall prove the existence of a minimizer of Φ on S. Since the maps E → E, u → u± are continuous but not differentiable, S is not a differentiable manifold even if N is one (as in Proposition 1.19). P ROOF OF T HEOREM 2.1. Clearly β = inf Φ(S)
inf Φ(N )
0 by Theorem 1.21. Let u ∈ S be a minimizer and suppose Φ ′ (u) = 0. As a consequence of (f1 ), for any v ∈ N , the function R+ ∋ t → Φ(tv) ∈ R achieves its unique maximum at t = 1. Therefore



  Φ su+ + tu− = Φ su+ + Φ(tu− ) < Φ u+ + Φ(u− ) = Φ(u)

(2.2)

for (s, t) ∈ R2+ \ {(1, 1)}. By the continuity of Φ ′ there exist α, δ > 0 such that Φ ′ (v)
α for v ∈ U3δ (u). Setting g:D =



 1 3 1 3 × → E, , , 2 2 2 2

g(s, t) = su+ + tu− ,

0 αδ (2.2) implies β0 := max Φ ◦ g(∂D) < β. For ε := min{ β−β 2 , 8 } and S = Uδ (u), Lemma 1.22 yields a deformation η such that η1 := η(1, ·) : E → E satisfies: • η1 (v) = v if Φ(v)  β − 2ε, • η1 (Φ β+ε ∩ S) ⊂ Φ β−ε , • Φ(η1 (v))  Φ(v) for all v ∈ E. It follows that max Φ ◦ η1 ◦ g(D) < β and that h := η1 ◦ g = g on ∂D. We shall show that h(D) ∩ S = ∅, which gives the contradiction inf Φ|S < β. Consider the map

2

ψ :D → R ,

ψ(s, t) :=



  1 1 ′ Φ h(s, t)+ h(s, t)+ , Φ ′ h(s, t)− h(s, t)− s t



and observe that ψ(s, t) = (0, 0) is equivalent to h(s, t) ∈ S. For (s, t) ∈ ∂D we have

  ψ(s, t) = Φ ′ su+ u+ , Φ ′ (tu− )u−

because h = g on ∂D. This implies deg(ψ, D, (0, 0)) = 1, hence ψ(s, t) = (0, 0) for some (s, t) ∈ D.  R EMARK 2.2. (a) The proof of Theorem 2.1 is a two-dimensional version of the proof of Theorem 1.21. A minimizer of Φ on N + or N − is a local minimum of Φ on N , hence a critical point of mountain pass type. Though not being a manifold, S is a kind of codimension 1 subset of N , a co-dimension 2 subset of E. Hence a minimizer of Φ on S is a critical point of Morse index 2 (if nondegenerate).

24

T. Bartsch et al.

(b) In [17], Lemma 3.2, Bartsch and Weth prove that S ∩ H 2 (Ω) is a co-dimension 2 submanifold of E ∩ H 2 (Ω) if f satisfies the conditions (f0′ ) and (f1′ ). (c) In [67] Liu and Wang give a slightly different proof of the above result without using the deformation lemma and therefore requiring less smoothness of the functional. 2.2. Localizing critical points A basic idea for localizing critical points can be formulated in a very general setting. Let X be a topological space and ϕ : G ⊂ [0, ∞) × X → X be a continuous semiflow on X. Here G = {(t, u) ∈ [0, ∞) × X: 0  t < T (u)} is an open subset of [0, ∞) × X, where T (u) ∈ (0, ∞] is the maximal existence time of the trajectory t → ϕ(t, u). We often write ϕ t (u) = ϕ(t, u). Given B ⊂ A ⊂ X we call B positively invariant in A if, for u ∈ B and T > 0 with ϕ t (u) ∈ A, 0  t  T , it follows that ϕ T (u) ∈ B. If even ϕ T (u) ∈ int B then B is said to be strictly positively invariant. In the case A = X, we simply call B (strictly) positively invariant. The notion of invariant sets has also been exploited in [65]. Recall that a continuous map γ : (C, D) → (A, B) between pairs D ⊂ C, B ⊂ A of topological spaces is nullhomotopic if there exists a homotopy H : (C × [0, 1], D × [0, 1]) → (A, B) with H (x, 0) = γ (x) and H (x, 1) ∈ B for all x ∈ C. L EMMA 2.3. Let A ⊂ X be positively invariant, B ⊂ A strictly positively invariant. Let f : (C, D) → (A, B) be not nullhomotopic, C a metric ( paracompact) space. Then 

  A(B) := x ∈ C  ∃t
0: ϕ t f (x) ∈ B = C.

P ROOF. We argue by contradiction. If A(B) = C then for each x ∈ C there exists τ (x)
0 with ϕ τ (x) (f (x)) ∈ intA B. Choose a neighborhood Vx of x in C with

 ϕ τ (x) f (y) ∈ intA B for all y ∈ Vx .

Let (πj )j ∈J be a partition of unity  subordinated to Vx : supp πj ⊂ Vxj . Now we define σ : C → [0, ∞), σ (x) := j ∈J πj (x)τ (xj ), and H : C × I → A, H (x, f ) := ϕ tσ (x) (f (x)). This homotopy shows that f is nullhomotopic. 

There also exists an equivariant version of Lemma 2.3 when a group G acts on A and C. The extension is straightforward and therefore omitted. In a typical application, ϕ is the negative gradient flow of a functional Φ : X → R, and B contains a sublevel set Φ b = {u ∈ X: J (u)  b}. A and B are closed and one wants to find a critical point in A \ B. For x ∈ C \ A(B) and u := f (x), one then has

 ϕ t (u) ∈ / B, hence Φ ϕ t (u) > b for all t. Then c := limt→∞ Φ(ϕ t (u))
b and there exists a sequence tn → ∞ with  



 ∇Φ ϕ tn (u) 2 = d J ϕ t (u)  → 0.  dt t=tn

The Dirichlet problem for superlinear elliptic equations

25

/ B because Thus un := ϕ tn (u), n ∈ N, is a (PS)c -sequence in A \ B. If un → u¯ then u¯ ∈ ¯ ∈ intA B, hence φ tn (u) ∈ intA B for some tn , a contradiction. Thus otherwise u¯ = ϕ t (u)  \ B. u¯ ∈ A We shall now present examples of strictly positively invariant sets which can be used to find nodal solutions of (2.1). Thus we are interested in finding critical points of the associated functional Φ outside of P ± = {u ∈ E: ±u
0 a.e.}.

(2.3)

Since P + and P − have empty interior one cannot use B = P + ∪ P − . E XAMPLE 2.4. Suppose Ω is bounded, and (f0′ ) and (f4 ) there exists m > 0 so that t → f (x, t) + mt is strictly increasing for all x ∈ Ω hold. We take u, vm := ∇u, ∇vL2 + mu, vL2 as scalar product in E and write  · m for the corresponding norm. Setting g(x, u) = t f (x, u) + mu and G(x, t) = 0 g(x, s) ds we can write Φ as 1 Φ(u) = u2m − 2



G(x, u) dx.

Ω

Thus the gradient of Φ with respect to the above scalar product has the form ∇Φ = Id −K

 with K(u) = (−
+ m)−1 g(·, u) .

 ∩ E invariant. The cones P ± := X ∩ P ± have nonBy (f0′ ), K leaves X = C 1 (Ω) X empty interior in X because Ω is bounded. We write u  v if v − u ∈ PX+ , and u ≪ v if v − u ∈ int(PX+ ). As a consequence of the strong maximum principle, K is strictly order preserving, that is, u 0 the vector field −∇Φ points at u + v into u + int PX+ , u + v − ∇Φ(u + v) = K(u + v) ≫ K(u).

26

T. Bartsch et al.

Similarly, if u is a supersolution then u + (PX± \ {0}) is strictly positively invariant. In order to find nodal solutions above some level α > 0 one can work with B := Φ α ∪ PX+ ∪ PX− which is strictly positively invariant for ϕ if there are no nodal solutions at the level α. E XAMPLE 2.5. Suppose Ω is bounded, and (f0′ ) and (f5 ) inft=0 f (x, t)/t > −∞ hold. Then we choose m > 0 so that f (x, t) + mt > 0 for all t > 0, f (x, t) + mt < 0 for all t < 0, and define ·, ·m , X and the order relation ≪ as in Example 2.4. The gradient of Φ with respect to this metric is not necessarily order preserving but it does satisfy u>0

⇒

K(u) ≫ 0 and u < 0

⇒

K(u) ≪ 0.

It follows as in Example 2.4 that the cones PX± \ {0} are strictly positively invariant. If Ω is unbounded or if f is only a Carathéodory function, the approach presented in  ∩ E. Either the Example 2.4 does not work because then one cannot work in X = C 1 (Ω) ± cones PX have empty interior or there is no flow on X due to a lack of regularity. Here one can often replace P ± by their open neighborhoods in E. D EFINITION 2.6. Let K : E → E be a continuous operator on a Banach space E. A set C ⊂ E is said to be K-attractive if there exists ε0 > 0 so that K(clos(Uε (C))) ⊂ Uε (C) = {u ∈ E: dist(u, C) < ε} for 0 < ε < ε0 . L EMMA 2.7. Let E be a Banach space, Φ ∈ C 1 (E) and C = C1 ∪ · · · ∪ Cn ⊂ E be a finite union of convex sets. Suppose ∇Φ = Id −K, and each Cj is K-attractive. Then Φ has no critical points in clos(Uε0 (C)) \ C, where ε0 is from Definition 2.6. Given ε ∈ (0, ε0 ] there exists a pseudo-gradient vector field V : E \ Fix(K) → E for Φ so that clos(Uε (C)) is strictly positively invariant for the flow associated to −V . P ROOF. For ε ∈ (0, ε0 ] and u ∈ ∂Uε (Cj ), we have K(u) ∈ Uε (Cj ). Since C1 , . . . , Cn are convex, a standard partition of unity argument yields a locally Lipschitz continuous map  \ Fix(K) → E such that K and

 ∈ Uε (Cj ) K(u)

for u ∈ ∂Uε (Cj ), j = 1, . . . , n,

    u − K(u)    2Φ ′ (u),

for u ∈ E \ Fix(K).

 1  ′ 2  Φ ′ (u) u − K(u)
Φ (u) , 2



Here are two examples of K-attractive sets. E XAMPLE 2.8. Suppose f satisfies (f0 ) and (f5 ). Thus there exists m > 0 as in Example 2.5, and we define K(u) = (−
+ m)−1 (f (·, u) + mu). If lim supt→0 |f (x, t)|/|t|
N/2. Then u is continuous by elliptic regularity. The last statement has been proved in [73], Lemma 1.  Now we suppose that (f0 ) and (f1 ) hold. Let N be the Nehari manifold and S ⊂ N the nodal Nehari set. P ROPOSITION 2.12. Suppose (f0 ) and (f1 ) hold. Let u be a critical point of Φ and fix n ∈ N. (a) If Φ(u)  inf Φ(S) + n · inf Φ(N ) then nod(u)  n + 1. (b) If Φ(u)  infv1 ,...,vn sup Φ(C(v1 , . . . , vn )) then nod(u)  n. The infimum extends over all n-tuples of linearly independent elements v1 , . . . , vn ∈ E, and  C(v1 , . . . , vn ) := { ni=1 λi · vi : λ1 , . . . , λn
0}. (c) If f is odd and Φ(u)  infV ⊂E,dim(V )=n sup Φ(V ) then nod(u)  n. Here the infimum extends over all n-dimensional linear subspaces of E.

28

T. Bartsch et al.

P ROOF. (a) Let u have k nodal domains Ω1 , . . . , Ωk such that v1 := u · χΩ1 > 0 and v2 := u · χΩ2 < 0. Clearly, vj := u · χΩj ∈ N and v1 + v2 ∈ S. Theorem 1.21 implies that vj is not a minimizer of Φ on N , hence Φ(u) = Φ(v1 + v2 ) +

k  j =3

Φ(vj ) > inf Φ(S) + (k − 2) inf Φ(N ).

This implies k  n + 1. (b) and (c) Suppose nod(u) > n and let Ω1 , . . . , Ωn be nodal domains of u. Then vi := u · χΩi ∈ N and Φ(vi ) = maxλ
0 Φ(λvi ). This implies

 Φ(u) > Φ(v1 + · · · + vn )
inf sup Φ C(v1 , . . . , vn ) . v1 ,...,vn

If f is odd then Φ(vi ) = maxλ∈R Φ(λvi ), hence

Φ(u) > Φ(v1 + · · · + vn )
inf sup Φ(V ). V



If Φ ∈ C 2 (E) then the Morse index provides a lower bound for the number of nodal domains. T HEOREM 2.13. Suppose (f0′ ) and (f1′ ) hold and let u be a critical point of Φ with Morse index µ(u). Then nod(u)  µ(u). P ROOF. Let Ω1 , . . . , Ωn be nodal domains of u and set vj := u · χΩj ∈ N . Then Φ ′′ (u)[vj , vj ] =
0, then (2.1) has a nodal solution.

The Dirichlet problem for superlinear elliptic equations

29

P ROOF. We first observe that (f0′ ) and (f2 ) imply (f5 ), so there exists m > 0 with f (x, t)t + mt 2 > 0 for t = 0, x ∈ Ω. Proceeding as in Example 2.5 we consider the scalar product u, vm := ∇u, ∇vL2 + mu, vL2 on E = H01 (Ω) and let ϕ t be the corresponding negative gradient flow of the energy functional associated to (2.1).  By (f2 ), there exist Let e1 > 0, e2 be linearly independent elements of X = E ∩ C 1 (Ω). R > 0 so that Φ(u)  0 for u ∈ span{e1 , e2 }, u
R. We define  C := u = se1 + te2 : s ∈ [−R, R], t ∈ [0, R]

and D := ∂C ⊂ span{e1 , e2 }. Let E1 ⊂ E be the first eigenspace of −
+ a(x). Since λ2 (−
+ a(x)) > 0, there exist α, ρ > 0 so that Φ(u)
α

for u ∈ E1⊥ , u = ρ.

Setting S := {u ∈ E1⊥ : u = ρ} the inclusion (C, D) ֒→ (E, E \ S) is well defined and not nullhomotopic. We fix β ∈ (0, α) so that (2.1) has no nodal solutions on the level β. If no such β exists the theorem is proved anyway. As stated in Example 2.4, the flow ϕ t leaves X invariant and B := (X ∩ Φ β ) ∪ PX+ ∪ PX− strictly positively invariant. By construction we have B ∩ S = ∅ which implies that the inclusion (C, D) ֒→ (X, B) is not nullhomotopic. Now Lemma 2.3 yields u ∈ C so that ϕ t (u) ∈ / B for every t > 0. It follows that there exists a sequence tn → ∞ with ∇Φ(ϕ tn (u))2m = | dtd Φ(ϕ tn (u))| → 0 as n → ∞. Thus un := ϕ tn (u), n ∈ N, is a (PS)c -sequence for some c
β, hence un → u¯ in E along a subsequence. By the ω-limit lemma from [10] un → u¯ in X. This implies u¯ ∈ /B because u¯ is a critical point of Φ, un ∈ / B, and B is strictly positively invariant. Thus u¯ is a nodal solution of (2.1).  T HEOREM 2.15. Suppose (f0′ ) with p < 2∗ and (f2 ) hold. If 0 ∈ / σ (−
+ a(x)) then (2.1) has a nodal solution. This theorem is a simple consequence of basic Morse theory which we recall here. T HEOREM 2.16. Let X be a Banach space, Φ ∈ C 1 (X, R), and ϕ t a (local) flow on X with the properties: (GRAD) If Φ ′ (u) = 0 then t → Φ(ϕ t (u)) is strictly decreasing. (COMP) If Φ(ϕ t (u)) → c ∈ R as t → ∞ then {ϕ t (u): t
0} is relatively compact. Let H∗ be any homology theory and let Ck (Φ, u) := Hk (Φ c , Φ c \ {u}), c = Φ(u), be the kth critical group of Φ at u. Finally, let B ⊂ X be strictly positively invariant such that Φ has only finitely many critical points u1 , . . . , uk ∈ X \ B. Setting Pk := rank Hk (X, B) and βk (ui ) := rank Ck (Φ, ui ), there exists Q ∈ N0 [[t]] such that  n ∞   k=0 j =1



βk (ui ) t k =

∞  k=0

Pk t k + (1 + t)Q(t).

30

T. Bartsch et al.

The proof is a variation of standard arguments, cf. [41], Theorem I.4.3. P ROOF OF T HEOREM 2.15. We consider Φ, ϕ t and X as in the proof of Theorem 2.14 and recall that the conditions (GRAD) and (COMP) hold. By Theorem 2.14, we may assume λ1 (−
+ a(x)) < 0. Then αe1 > 0 is a strict supersolution of (2.1) for α > 0 small. Here e1 > 0 is a first Dirichlet eigenfunction of −
+ a(x). By Example 2.4, A := (αe1 + PX+ ) ∪ (−αe1 + PX− ) is strictly positively invariant. Condition (f2 ) implies that Hk (X, Φ −β ) = {0} for β > 0 large (see [41], Lemma III.2.3), any k ∈ Z. (f2 ) also implies Φ(tu) → −∞ as t → ∞ for u = 0. It is now easy to show that B := A ∪ Φ −β can be deformed into Φ −β by a deformation of the form (t, u) → tu for u ∈ B, 0  t  T (u). As a consequence Hk (X, B) = {0}, k ∈ Z. Since B is strictly positively invariant, we can apply Theorem 2.16. By assumption, 0 ∈ X \ B is nondegenerate, hence Cµ (Φ, 0) = {0}, where µ is the Morse index of 0. It follows that there exists a critical point u¯ ∈ X \ B, u¯ = 0, with Cµ+1 (Φ, u) ¯ = {0} or Cµ−1 (Φ, u) ¯ = {0}. Clearly u¯ is a nodal solution of (2.1).  Next we assume (f1 ), that is, t →  f (x, t)/|t| is strictly increasing on R+ and on R− . Recall the nodal Nehari set


  ± ± ± ∇u∇u = f (x, u)u S = u ∈ E: u = 0, Ω

Ω

from Section 2.1. T HEOREM 2.17. Suppose (f0′ ) with p < 2∗ , (f1′ ) and (f2 ) hold. If λ2 (−
+ a(x)) > 0 then inf Φ(S) > 0 and it is achieved by a nodal solution with precisely two nodal domains and Morse index 2. A minimizer of Φ|S is a least energy nodal solution. Geometric properties of these on radial domains will be investigated in Section 2.5. P ROOF OF T HEOREM 2.17. Consider a minimizing sequence en ∈ S, n ∈ N, of Φ|S , and define  Cn := u = sen+ + ten− : s ∈ [−Rn , Rn ], t ∈ [0, Rn ] ,

where Rn > 0 is chosen so that Φ(u)  0 for u ∈ span{en+ , en− }, u
Rn . Observe that Φ(en ) = max Φ(Cn ) as in (2.2). Now, the same argument as in the proof of Theorem 2.14 yields a nodal solution un ∈ S with Φ(un )  Φ(en ). Moreover, Φ(un ) is bounded away from 0. In fact, Φ(un )
inf Φ({u ∈ E1⊥ : u = ρ}) > 0, where E1 is the first eigenspace of −
+ a(x) and ρ > 0 is small. Thus (un )n is a minimizing sequence of Φ|S consisting of critical points. Thus inf Φ|S is achieved by the Palais–Smale condition and because S is closed in E. The equality nod(u) ¯ = 2 for a minimizer u¯ follows from Proposition 2.12. By Remark 2.2(b), S ∩ H 2 (Ω) is a submanifold of E ∩ H 2 (Ω) with co-dimension 2. The Hessian Φ ′′ (u) ¯ of Φ at a minimizer u¯ of Φ|S is therefore positive semidefinite on the tangent space V := Tu¯ (S ∩ H 2 (Ω)) ⊂ E ∩ H 2 (Ω) which has co-dimension 2 in

The Dirichlet problem for superlinear elliptic equations

31

E ⊂ E which has coE ∩ H 2 (Ω). It follows that Φ ′′ (u) is positive semidefinite on V dimension 2 in E.  R EMARK 2.18. (a) The first existence results for one nodal solution under similar hypotheses as those considered here are due to [15,36,50]. Improved versions and related results can be found in [17,37,45]. In [90] of Wang, besides the positive and negative solutions given in [5] by Ambrosetti and Rabinowitz a third solution was found by two different methods: linking and the Morse theory. No nodal information was given for the third solution in [90], but the methods used there seem to be suggestive for the nodal nature of the solution, motivating some of the results mentioned above. Using symmetries of the domain, the existence of nodal solutions can be proved more easily. Results of this type abound, in particular, for radially symmetric domains. (b) In the recent paper [67] of Liu and Wang, it is proved that inf Φ(S) > 0 is achieved by a nodal solution with precisely two nodal domains provided (f0 ), (f1 ) and the condition (f3 ), that is, lim|t|→∞ F (x, t)/t 2 = ∞ uniformly in x ∈ Ω, hold. Observe that this condition is weaker than the classical Ambrosetti–Rabinowitz condition (f2 ). Under the same conditions a positive and a negative solutions are given as in Theorem 1.23. The proofs is in the same spirit of that for Theorem 1.23. 2.5. Geometric properties of least energy nodal solutions on radial domains We consider




−
u + a |x| u = f |x|, u u=0

in Ω, on ∂Ω,

(2.4)

when Ω is a ball or an annulus and a and f are radial. It is then natural to ask about the symmetry of solutions of (2.4). We shall address this question for the least energy nodal solution when it exists. T HEOREM 2.19. Suppose Ω is a ball or an annulus, a ∈ L∞ (Ω) and f are radial, and f satisfies (f0 ) and (f1 ). Suppose, moreover, that f is Hölder continuous on Ω × [−R, R] for every R > 0. Then every least energy nodal solution is foliated Schwarz symmetric. This result is due to [19]. Its proof is based on an elementary symmetrization, the polarization, and the maximum principle as the one of Theorem 1.24. If f = f (u) is independent of x the problem whether a least energy nodal solution is radial has been settled recently by Aftalion and Pacella [1]. The remaining part of this section contains several results from [1]. We begin with the following observation relating the number of nodal domains to the Morse index. T HEOREM 2.20. Suppose Ω is a ball or an annulus, and f ∈ C 1 (R) satisfies f (0)
0. Let u be a radial solution of
−
u = f (u) in Ω, u=0 on ∂Ω.

32

T. Bartsch et al.

Then the Morse index of u is at least (N + 1)(nod(u) − 1). P ROOF. Let Ω = A(r, R) = int{x ∈ RN : r  |x|  R} and let r = r0 < r1 < · · · < rk−1 < rk = R be such that Ai = A(ri−1 , ri ), i = 1, . . . , k = nod(u), are the nodal domains of u. We consider the domains Bi = A(ri−1 , ri+1 ) and Bij = {x ∈ Bi : xj < 0}, i = 1, . . . , k − 1, j = 1, . . . , N . Let µij be the first Dirichlet eigenvalue of −
− f ′ (u) in Bij and ψij > 0, a corresponding eigenfunction. We claim that µij < 0. If this has been proved let vij ∈ H01 (Bi ) ⊂ H01 (Ω) be the extension of ψij which is odd in xj . Then vij is a Dirichlet eigenfunction of −
− f ′ (u) in Bi with eigenvalue µij < 0. Since vij changes sign, there exists a positive eigenfunction vi0 ∈ H01 (Bi ) ⊂ H01 (Ω) of −
− f ′ (u) with eigenvalue µi0 < µij < 0. It follows that the quadratic form v → (−
− f ′ (u))v, vL2 is negative on span{vij : i = 1, . . . , k − 1, j = 0, . . . , N}. Since the vij , i = 1, . . . , k − 1, j = 0, . . . , N , are linearly independent by construction, the negative eigenspace of −
− f ′ (u) in H01 (Ω) has dimension at least (k − 1)(N + 1) = (nod(u) − 1)(N + 1). It remains to prove µij < 0. Observe that uj := ∂u/∂xj solves −
uj = f ′ (u)uj in Ω. Let ∂o Bi = {x ∈ RN : |x| = ri+1 } be the outer boundary of Bi . Since u is radial, uj (x) = 0 if xj = 0. Suppose u > 0 on A(ri , ri+1 ) ⊂ Bi and u < 0 on A(ri−1 , ri ) ⊂ Bi . Then uj (x) > 0 for x ∈ ∂Bi with xj < 0, and Cij := {x ⊂ Bij : uj (x) < 0}
Bij is a nonempty open subset of Bij with ∂Cij ⊂ Bij ∪ {x: xj = 0}. It follows that uj = 0 on ∂Cij , hence uj ∈ H01 (Cij ) ⊂ H01 (Bij ). Therefore µij =
uk is a supersolution of (2.1) then u
0. Here k0 = µ(−
+ a(x)) + dim ker(−
+ a(x)) + 1, where µ(−
+ a(x)) denotes the Morse index. The existence of infinitely many solutions as well as property (a) is a classical result of Ambrosetti and Rabinowitz [5] based on the symmetric mountain pass theorem. The fact that the solutions are nodal and properties (b)–(d) has been proved in [10,17] where co-homological linking properties are being used; see also [61], where the existence of nodal solutions is proved using the Lusternik–Schnirelmann theory. It is not known whether nod(uk ) → ∞ as k → ∞. In fact, it is not even known whether nod(uk )
3 for some k. In the following, we sketch the proof from [61] for the existence of infinitely many nodal solutions using a variant of the symmetric mountain pass theorem.

34

T. Bartsch et al.

P ROOF  OF T HEOREM 2.24.  Set λi = λi (−
+ a(x)), Ei = ker(−
+ a(x) − λi ), Yk = ki=1 Ei and Zk = ∞ i=k Ei . Then it is easy to check that, for all k with λk > 0, there are ρk > rk > 0 such that supYk ∩Bρ (0) Φ(u)  ak := 0 and bk := infZk ∩∂Brk (0) Φ(u) > 0 k  ∩ E. Then Yk ⊂ X for any k
1. Let P ± be as in Examand bk → ∞. Let X = C 1 (Ω) X ple 2.4. Then PX± are invariant sets of the negative gradient flow. Moreover, for k
2,  ± \ {0}, Zk ∩ PX± = {0}. uφ dx = 0 while This can be seen by noting that, for all u ∈ P 1 Ω  for u ∈ Zk , Ω uφ1 dx = 0, where φ1 is the first eigenfunction of the Laplacian operator on Ω. As a consequence of the Borsuk–Ulam theorem, the inclusion (Bk , ∂Bk ) ֒→ (E, Φ 0 ∪ P + ∪ P − ) is not nullhomotopic; here Bk = {u ∈ Yk : u  ρk }. Lemma 2.3 or the deformation lemma from [61] or the proof of [61], Theorem 3.2, yield a critical point outside of P + ∪ P − . In fact, it follows that ck := inf

h∈Γk

sup u∈h(Bk )\(PX+ ∪PX− )

Φ(u)
bk

is a critical value with a critical point in E \ (P + ∪ P − ); here Γk = {h ∈ C(Bk , X): h is odd, and h(u) = u if u = ρk }.  The idea in the above proof can be adapted for many other boundary value problems with different nonlinearity; see [61] for more examples. See also [92] for an abstract version of a variational principle in the presence of invariant sets of the flows. Nodal solutions for nonlinear eigenvalue problems with superlinear nonlinearity have been studied in [47, 61,63]. Without symmetry of the domain or the nonlinearity it is not known whether the superlinear problems considered here have “many”, even infinitely many solutions. Now we discuss the role of the domain. If Ω is a radial domain and f = f (|x|, u) is radially symmetric in x then one can look for radially symmetric solutions u(x) = v(|x|). Rewriting (2.1) as an ordinary differential equation for v(r), ODE methods can be applied. In this setting there exist solutions with any prescribed number of nodal domains for a large class of superlinear nonlinearities. The following recent result gives a lower bound on the number of solutions when the domain is large in the sense that it contains balls of large sizes. T HEOREM 2.25. Suppose f = f (u) is independent of x ∈ Ω and satisfies (f0′ ) with p < 2∗ , (f1′ ), (f2 ), and f ′ (0) < 0. Then there exists R > 0 such that if Ω contains a ball of radius R then (2.1) has at least 3 nodal solutions u1 , u2 , u3 with nod(u1 ) = nod(u2 ) = 2, 2  nod(u3 )  3. The proof can be found in [18]. This result is in some sense of the singularly perturbed type; cf. Remark 1.26. It yields 3 nodal solutions of the singularly perturbed problem (1.15) on any domain. As mentioned in Remark 1.26 there is a huge literature about singularly perturbed problems, which we cannot present in this chapter.

The Dirichlet problem for superlinear elliptic equations

35

2.7. Perturbed symmetric functionals It is natural to ask whether the solutions continue to exist if an odd function is perturbed, e.g., f (x, u) = |u|p−2 u + g(x). This subsection is devoted to semilinear elliptic problems of the type


−
u = |u|p−2 u + h(x) in Ω, u=0 on ∂Ω,

(2.6)

where Ω ⊂ RN is a smooth and bounded domain, 2 < p < 2∗ and h ∈ L2 (Ω) is fixed. Solutions of (2.6) are critical points of 1 Φ(u) = 2



1 |∇u| dx − p Ω 2



p

Ω

|u| dx −



Ω

hu dx,

u ∈ H01 (Ω).

When h = 0 and p < 2∗ the corresponding functional 1 Φ0 (u) = 2



1 |∇u| dx − p Ω 2



|u|p dx,

u ∈ H01 (Ω),

is odd and has infinitely many critical points. The following result is due to Bahri and Lions [9] improving earlier work of Bahri and Berestycki [7] and, independently, Struwe [86]. T HEOREM 2.26. If 2 < p < (2N − 2)/(N − 2) then, for every h ∈ L2 (Ω), problem (2.6) has infinitely many solutions. The proof is based on the idea that Φ is a kind of perturbation of Φ0 . The symmetric linking which is based on the Borsuk–Ulam theorem and used in the proof of Theorem 2.24 applies to Φ0 and yields a nonsymmetric linking for Φ. A different approach to perturbed symmetric functions is due to Bolle; see [27,28]. It is still not known whether (2.6) has infinitely many solutions for any 2 < p < 2∗ and any h. However, Bahri proved in [6] the following generic existence result for the full range of p. T HEOREM 2.27. If 2 < p < 2∗ then the set of h ∈ H −1 (Ω) such that problem (2.6) has infinitely many solutions is a dense residual set in H −1 (Ω).

3. Problems on the entire space In this section we consider the problem


−
u + V (x)u = f (x, u), u(x) → 0

x ∈ RN , as |x| → ∞.

(3.1)

36

T. Bartsch et al.

Depending upon the potential function V (x) we review the existence of signed solutions, nodal solutions and multiplicity of solutions as well as their qualitative properties in some cases. Let   1 2 2 Φ : E → R, Φ(u) = |∇u| + V (x)u dx − F (x, u) dx 2 RN RN t be the functional associated to (3.1). As usual, F (x, t) = 0 f (x, s) ds. The nonlinearity f will always satisfy (f0 ) and is subcritical, so that Φ is defined on the space




 E = u ∈ H 1 RN : V (x)u2 dx < ∞ RN

and is of class C 1 . If V is bounded then E = H 1 (RN ) and the Palais–Smale condition does not hold. On the contrary, if V (x) → ∞ as |x| → ∞ then Φ does satisfy the Palais– Smale condition. This compact case is most closely related to the bounded domain case in terms of the results and methods and will be dealt with in Section 3.1. If V and f are radially symmetric it is natural to look for radially symmetric solutions. Constraining E to the space Erad of radial functions u = u(|x|) compactness is recovered. We deal with the symmetric case in Section 3.2. In fact, in both Sections 3.1 and 3.2 weaker conditions than those just mentioned will be considered. In the remaining subsections we deal with bounded, nonradial potentials, in particular with potentials which are periodic in the xi -variables. Since f is subcritical this case may be called locally compact, a notion going back to Lions’ seminal work [64]. Locally the problems have compactness, and the compactness is only lost from the mass going to infinity. Lions’ concentration–compactness principle is an important tool in dealing with this and will be used. In Section 3.3 we consider potentials having a potential well whose steepness is controlled by a parameter. In Section 3.4 we present two results on the existence of a ground state solution for bounded potentials. In the periodic case, one can construct bound states having multibumps. The first result of this type is due to Coti Zelati and Rabinowitz [48] who constructed positive multibump solutions. Here we present the modified approach from [68] which can also be used to obtain nodal multibump-type solutions and to control the number of nodal domains. So far we always assumed inf V > 0. Finally we discuss the case where V is periodic and negative somewhere. The periodicity implies that the spectrum of −
+ V is purely continuous and consists of a disjoint union of closed intervals. In Section 3.6 we deal with the case where −
+ V has essential spectrum below 0.

3.1. The compact case  We consider the space E = {u ∈ H 1 (RN ): RN V (x)u2 dx < ∞} together with the scalar product   ∇u · ∇v + V (x)uv. u, vE = RN

RN

The Dirichlet problem for superlinear elliptic equations

37

In general the embedding from E into Lp (RN ) is not compact, for example, when V is bounded so that E = H 1 (RN ). We present general conditions on V which guarantee that the embedding is compact. (V0 ) V ∈ C(RN , R) and infRN V (x) > 0. (V1 ) There exists r0 > 0 such that, for any M > 0,

  lim m x ∈ RN : |x − y|  r0 , V (x)  M = 0,

|y|→∞

where m denotes the Lebesgue measure in RN . Without loss of generality we assume infRN V (x) = 1. L EMMA 3.1. Under (V0 ) and (V1 ), the embedding from E into Lp (RN ) is compact for 2  p < 2∗ . P ROOF. Let (un ) be bounded in E and assume un → 0 weakly in E. We have to show that un → 0 in Lp (RN ) for 2  p < 2∗ . By the interpolation inequality we only need to consider p = 2. The Sobolev embedding theorem implies un → 0 in L2loc . Thus it suffices to  show that, for any ε > 0, there is R > 0 such that B c (0) u2n < ε; here BRc (0) = RN \ BR (0). R  Let (yi ) be a sequence of points in RN satisfying RN ⊂ ∞ i=1 Br0 (yi ) and such that each point x is contained in at most 2N such balls Br0 (yi ). Let AR,M = {x ∈ BRc | V (x)
M} and BR,M = {x ∈ BRc | V (x) < M}. Then 

AR,M

u2n



1  M

RN

V (x)u2n ,

and this can be made arbitrarily small by choosing M large. Choose q such that 2q  2∗ and let q ′ = q/(q − 1) be the dual exponent. Then for fixed M > 0, 

BR,M

u2n



∞  

i=1 BR,M ∩Br0 (yi )



∞   i=1

 εR

BR,M ∩Br0 (yi )

∞  

∞  

 CεR 2



RN

1/q

2q

i=1 Br0 (yi ) N

 |un |2q

BR,M ∩Br0 (yi )

i=1

 CεR

u2n

|un |

1/q ′ m BR,M ∩ Br0 (yi )

1/q

 |∇un |2 + u2n

 |∇un |2 + |un |2 ,

38

T. Bartsch et al. ′

where εR = supyi (m(BR,M ∩ Br0 (yi )))1/q . Assumption (V1 ) states that εR → 0 as R → ∞. Thus we may make this term small by choosing R large.  Concerning the nonlinearity f we recall the following conditions. (f0 ) f : RN × R → R is a Carathéodory function with f (x, t) = o(t) as t → 0. There exist C > 0 and p  2N/(N − 2) such that |f (x, t)|  C(|t| + |t|p−1 ) for all x ∈ RN , t ∈ R. (f1 ) For every x ∈ RN , the function R \ {0} → R, t → f (x, t)/|t|, is strictly increasing. (f2 ) There exists θ > 2 such that 0 < θ F (x, t)  f (x, t)t for all x ∈ RN , all t ∈ R. The first result is concerned with minimizers of Φ on the Nehari manifold  N = u ∈ E \ {0}: Φ ′ (u)u = 0 .

In fact, we find a least energy positive and a least energy negative solution, that is, we minimize Φ on N± = {u ∈ N : ±u
0}. T HEOREM 3.2. If (V0 ), (V1 ), (f0 ) with p < 2∗ , (f1 ) and (f2 ) hold then inf{Φ(u): u ∈ N± } is achieved, hence (3.1) has a least energy positive solution and a least energy negative solution. P ROOF. Using the compact embedding lemma (Lemma 3.1) the proof proceeds as the one of Theorem 1.23.  Next we consider the nodal Nehari set  S = u ∈ E: u+ ∈ N , u− ∈ N  = u ∈ E: u+ =  0 = u− , Φ ′ (u)u+ = 0 = Φ ′ (u)u−

and find a least energy nodal solution.

T HEOREM 3.3. Suppose (V0 ), (V1 ), (f0 ) with p < 2∗ , (f1 ) and (f2 ) hold. Then inf Φ|S is achieved by a nodal solution that has exactly two nodal domains. P ROOF. Using the compact embedding Lemma 3.1 the proof proceeds as the one of Theorem 2.17.  R EMARK 3.4. As commented in Remark 2.18, (f2 ) can be replaced by the weaker condition (f3 ). This is done in [67]. T HEOREM 3.5. Assume (V0 ), (V1 ), (f0 ) with p < 2∗ , (f1 ) and (f2 ). If f is odd in u, then (3.1) has an unbounded sequence of nodal solutions uk such that 2  nod(uk )  k. Details of the proofs of these and related results can be found in [12].

The Dirichlet problem for superlinear elliptic equations

39

 R EMARK 3.6. Under the stronger condition that RN V (x)−1 dx < ∞, a nonlinearity with a combination of convex and concave terms was considered recently by Liu and Wang in [69] in which multiplicity of nodal solutions was given.

3.2. The radially symmetric case We consider the following class of symmetric functions. We fix a decomposition N = N1 + · · · + Nk with Ni
2. For x ∈ RN , we write x = (x1 , . . . , xk ) with xi ∈ RNi and define

N 



 1 Hsym R = u ∈ H 1 RN : u(x) = u |x1 |, . . . , |xk | .

1 (RN ) consists precisely of the radial H 1 -functions. If k = 1 then Hsym 1 (RN ) into Lp (RN ) is compact for 2 < p < 2∗ . T HEOREM 3.7. The embedding of Hsym

For the proof we need the following lemma due to Lions [64]. L EMMA 3.8. Let (un ) be a bounded sequence in H 1 (RN ). Assume that there is r > 0 and 2  q  2∗ such that lim sup sup n→∞ y∈RN

Then limn→∞



RN



Br (y)

|un |q dx = 0.

|un |p dx = 0 for 2 < p < 2∗ . If q = 2∗ , p can be taken as 2∗ .

Now we give the proof of Theorem 3.7. 1 (RN ) which we may P ROOF OF T HEOREM 3.7. Let (un ) be a bounded sequence in Hsym 1 N assume converges weakly to 0 in Hsym (R ). We want to show that (un ) converges to 0 in Lp (RN ). By Lemma 3.8, if un does not converge to zero in Lp (RN ) there are r > 0, δ > 0 and yn ∈ RN such that

lim inf n→∞



Br (yn )

|un |p
δ.

Since the embedding H 1 (Ω) ֒→ Lp (Ω) is compact for bounded domains Ω, we have |yn | → ∞. Up to a subsequence we may assume that there is 1  i  k such that the ith components yn,i ∈ RNi of (yn ) go to infinity. Using the radial symmetry in RNi , as n → ∞ we and more disjoint balls of radius r on which the Lp norm of un is the  find more p 1 (RN ) and therefore same as Br (yn ) |un | , contradicting the fact that un is bounded in Hsym p N  in L (R ).

40

T. Bartsch et al.

(S) V (x) and f (x, u) are radially symmetric in each of the variables xi ∈ RNi , i = 1, . . . , k. 1 (RN ) is a critical point of Φ by If (S) holds then a critical point of Φ constrained to Hsym the principle of symmetric criticality. Thus, with the compactness Theorem 3.7 available we easily obtain existence results as in the last section. The proofs are very similar and we omit them. T HEOREM 3.9. Assume (V0 ), (f0 ) with p < 2∗ , (f1 ), (f2 ) and (S). Then the infimum of Φ 1 (RN ) is achieved, hence (3.1) has a least energy positive solution and a on N ± ∩ Hsym 1 (RN ). least energy negative solution in Hsym T HEOREM 3.10. Assume (V0 ), (f0 ) with p < 2∗ , (f1 ), (f2 ) and (S). Then (3.1) has a nodal 1 (RN ) which has exactly two nodal domains. solution in Hsym T HEOREM 3.11. Assume (V0 ), (f0 ) with p < 2∗ , (f1 ), (f2 ) and (S). If f is odd in u, 1 (RN ) with then (3.1) has an unbounded sequence of nodal solutions uk ∈ Hsym 2  nod(uk )  k. R EMARK 3.12. If the equation is autonomous then radial solutions have been obtained by Berestycki and Lions [25]. The autonomous case has been further investigated by many authors. For more recent results, see the paper by Jeanjean and Tanaka [55] and the references therein. Nonradial solutions have been obtained in dimensions N = 4 or N
6 by Bartsch and Willem [20] provided f is odd and V and f are radial functions of x. In dimension N = 5 a nonradial solution can be obtained combining the idea from [20] with the concentration–compactness method; see the paper [71] by Lorca and Ubilla. Whether or not a nonradial solution exists in dimension N = 3 is open. 3.3. The steep potential well case In this section, we consider potentials depending on a parameter that controls the depth of the potential well. More precisely, we want to find bound states of the equation (Sλ )

 −
u + λa(x) + 1 u = f (x, u)

in RN .

We require the following assumptions on the potential function Vλ (x) = λa(x) + 1. (a1 ) a : RN → R is continuous, a
0, Ω := int a −1 (0) = ∅, has smooth boundary and  = a −1 (0). Ω (a2 ) There exist M0 > 0 and r0 > 0 such that

  lim m x ∈ RN : |x − y|  r0 , a(x)  M0 < ∞,

|y|→∞

where m denotes the Lebesgue measure on RN .

The Dirichlet problem for superlinear elliptic equations

41

The set Ω in assumption (a1 ) is the bottom of the potential well. It is allowed that Ω is unbounded. It is also allowed that a is bounded. In view of the results and the proofs from the last sectionwe need to study the compactness of the problem first. On the space E = {u ∈ H 1 (RN ): RN a(x)u2 dx < ∞} we use the norms u2λ

=



RN

  |∇u|2 + λa(x) + 1 u2 dx

for λ
0,

which are equivalent norms on E. We shall occasionally write Eλ for the Hilbert space E equipped with the norm  · λ . The functional associated to (Sλ ) is given by  

 2 1 2 |∇u| + λ a(x) + 1 u dx − F (x, u) dx Φλ (u) = 2 RN RN  1 = u2λ − F (x, u) dx. (3.2) 2 RN If (f0 ) holds we have Φλ ∈ C 1 (Eλ , R) for any λ
0. We have the following parameterdependent compactness result for Φλ . P ROPOSITION 3.13. Suppose (a1 ), (a2 ), (f0 ) and (f2 ). For any C0 > 0 there exists Λ0 > 0 such that Φλ satisfies the (PS)c -condition for all λ
Λ0 and all c  C0 . Since the proof of Proposition 3.13 is rather technical we just refer the reader to [13,16]. The result should also be compared with Theorem 3.1 in which global compactness is given. T HEOREM 3.14. Assume (a1 ), (a2 ), (f0 ) with p < 2∗ and (f2 ). Then for λ large, (Sλ ) has a positive solution and a negative solution. P ROOF. If, in addition, (f1 ) holds then one obtains the positive and the negative solution by minimizing on the positive (resp. negative) Nehari manifold as before, using the compactness result from Proposition 3.13. If (f1 ) does not hold one may apply the mountain pass theorem instead of minimizing.  T HEOREM 3.15. Assume (a1 ), (a2 ), (f0 ) with p < 2∗ , (f1 ) and (f2 ). Then for λ large, (Sλ ) has a nodal solution which has exactly two nodal domains. P ROOF. Here we minimize on the nodal Nehari set associated to the problem and observe that at the minimal level the Palais–Smale condition holds for λ large by Proposition 3.13.  Details of the proofs of these two results can be found in [13,14]. T HEOREM 3.16. Assume (a1 ), (a2 ), (f0′ ) with p < 2∗ , (f1 ) and (f2 ). If f is odd in u, then, for any integer k, there is Λk such that for λ
Λk , equation (Sλ ) has k pairs of

42

T. Bartsch et al.

nodal solutions uj , j = 1, . . . , k, such that the number of nodal domains of uj is bounded by j + 1. For the existence of multiple nodal solutions we need to use a combination of a minimax procedure and invariant sets of the gradient flow as was done in [12]. We define Kλ (u) := (−
+ Vλ )−1 f (·, u), and consider the flow on E associated to 

d t t dt ηλ (u) = −ηλ (u) + Kλ ηλ0 (u) = u.

t  ηλ (u) ,

We claim that the cones P ± = {u ∈ E: ±u
0} are Kλ -attractive in the sense of Definition 2.6. This is in fact independent of λ. For any M ⊂ E and ε > 0, Mε denotes the closed ε-neighborhood of M, i.e.,  Mε := u ∈ E: distλ (u, M)  ε .

L EMMA 3.17. Assume (a1 ), (a2 ) and (f0 ). Then there exists ε0 > 0 such that

 

  Kλ P ± ε ⊂ int P ± ε for all 0 < ε  ε0 , λ
0,

so P ± is Kλ -attractive uniformly in λ. Consequently,

 

  ηλt P ± ε ⊂ int P ± ε for all t > 0, 0 < ε  ε0 and λ
0.

P ROOF. We write Vλ (x) = λa(x) + 1. For u ∈ E, we denote v = Kλ (u) and u+ = max{0, u}, u− = min{0, u}. Note that, for any u ∈ E and 2  p  2∗ , u− Lp = inf u − wLp .

(3.3)

w∈P +

Since v − 2λ = (v, v − )λ =



RN

(∇v · ∇v − + Vλ vv − ) dx =



RN

f (x, u)v − dx,

the fact that v + ∈ P + and v − v + = v − implies 

distλ v, P + · v − λ  v − 2λ 



RN

f (x, u− )v − dx.

As a consequence of (f0 ) there exists a δ > 0 and C1 > 0 such that ∗ −2

f (x, s)
(1 − δ)s + C1 |s|2

s

for s  0.

(3.4)

The Dirichlet problem for superlinear elliptic equations

43

Thus 

RN



f (x, u− )v − dx 

RN

 ∗ (1 − δ)u− + C1 |u− |2 −2 u− v − dx ∗

2 −1 − ∗  (1 − δ)u− L2 v − L2 + C1 u− L 2∗ v L2 .

From the Sobolev embedding and (3.3)–(3.5),

 distλ v, P + · v − λ

(3.5)

∗

−  (1 − δ) inf u − wL2 v − λ + C2 inf u − w2L2−1 ∗ v λ , w∈P +

w∈P +

which implies (if v − λ = 0)

 2∗ −1 distλ v, P +  (1 − δ) inf u − wL2 + C2 inf u − wL 2∗ w∈P +

w∈P +

∗ −1

 (1 − δ) inf u − wλ + C3 inf u − w2λ w∈P +

w∈P +



2∗ −1 = (1 − δ) distλ u, P + + C3 distλ u, P + .

Therefore, there exists ε0 > 0 such that if distλ (u, P + )  ε0 then



 distλ v, P + < distλ u, P + .



The proof of Theorem 3.16 is based on Lusternik–Schnirelmann methods similar to those used in the proof of Theorem 2.24 and the above compactness result.

R EMARK 3.18. Theorem 3.16 still holds when one replaces the potential λa + 1 by λa + a0 with a0 ∈ L∞ (RN ) and such that −
+ a0 is invertible. In this case 0 is a saddle point instead of a local minimum of the functional. The invariant sets given in Lemma 3.17 do not work, but one can use the invariant sets constructed in [66] in this case. 3.4. Ground state solutions for bounded potentials In this section we consider bounded potentials V which cause some compactness problems. We shall present two results on the existence of a ground state solution. In order to keep the presentation simple we only deal with the case of a homogeneous nonlinearity. Thus we want to find a ground state solution u ∈ E = H 1 (RN ) of the equation −
u + V (x)u = |u|p−2 u, where 2 < p < 2∗ . We shall first treat a periodic potential and assume

(3.6)

44

T. Bartsch et al.

(V3 ) V is 1-periodic in each of its variables. The case where V is Ti -periodic in xi , i = 1, . . . , k, can be easily reduced to the period 1 case. T HEOREM 3.19. Assume (V0 ) and (V3 ). Then (3.6) has a least energy positive solution. P ROOF. We consider the following minimization problem on E = H 1 (RN ):   m := inf |∇u|2 + V (x)u2 dx, M = u ∈ E: uLp = 1 . u∈M RN

The Sobolev inequality implies m > 0. A minimizer of m gives a solution of (3.6) by a simple scaling factor. We show next that m is achieved. Let (un ) be a minimizing sequence for m. Then up to a subsequence un → u weakly in E. Since un Lp = 1, by Lemma 3.8 there is r > 0, δ > 0 and (yn ) ⊂ RN such that  lim inf |un |p
δ. n→∞

Br (yn )

By increasing r we may assume that all yn have integer coordinates. Setting vn (x) = un (x + yn ) we see that (vn ) is also a minimizing sequence for m because V is periodic. After passing to a subsequence we may assume that vn → v weakly. It follows from  lim inf |vn |p
δ n→∞

Br (0)

that v = 0. If vLp = 1 we are done. If vLp < 1 we produce a contradiction as follows. By the Brezis–Lieb lemma we have p

p

1 = vn − vLp + vLp + o(1)

as n → ∞.

p

p

Set a = vLp so that vn − vLp → 1 − a as n → ∞. m + o(1) = =



RN



RN

|∇vn |2 + V (x)|vn |2 2

2

|∇v| + V (x)|v| +



RN

  ∇(vn − v)2 + V (x)|vn − v|2 + o(1)


m v2Lp + vn − v2Lp + o(1).

Passing to the limit we get m
m(a 2/p + (1 − a)2/p ), a contradiction. Now we consider nonperiodic potentials satisfying (V4 ) lim|x|→∞ V (x) = V∞ := supRN V (x) < ∞. T HEOREM 3.20. If (V0 ) and (V4 ) hold then (3.6) has a least energy positive solution.



The Dirichlet problem for superlinear elliptic equations

45

P ROOF. We consider again the minimization problem on E = H 1 (RN ),  m := inf |∇u|2 + V (x)u2 dx u∈M RN

and compare it with m∞ := inf



u∈M RN

|∇u|2 + V∞ u2 dx,

where M = {u ∈ E: uLp = 1} is as before. The Sobolev inequality yields m > 0, and it is easy to see that a minimizer of m gives a solution of (3.6). If V (x) ≡ V∞ we may use the last theorem to get the existence of a positive solution. Thus we may assume V (x) < V∞ somewhere. Then using the minimizer of m∞ as a test function we obtain m < m∞ . We show next that m is in fact achieved on M. Let (un ) be a minimizing sequence for m. Then up to a subsequence un → u weakly in E. Since un Lp = 1, by Lemma 3.8 there exist r > 0, δ > 0 and (yn ) ⊂ RN such that  lim inf |un |p
δ. n→∞

Br (yn )

We claim that there exists a bounded sequence (yn ). If this is not true, we have u = 0 and q un → 0 in Lloc (RN ). Then  m + o(1) = |∇un |2 + V (x)|un |2 dx RN

= =





 V (x) − V∞ |un |2 dx

RN

|∇un |2 + V∞ |un |2 dx +

RN

|∇un |2 + V∞ |un |2 dx + o(1)



RN


m∞ + o(1). Passing to the limit we have a contradiction with m < m∞ . Thus we proved that (yn ) is bounded which implies u = 0. If uLp = 1 we are finished. If uLp < 1 we can produce a contradiction as in the proof of the last theorem.  3.5. More on periodic potentials In this subsection we describe the multibump type solutions for the periodic nonlinear Schrödinger equation constructed by Coti Zelati and Rabinowitz in [48] and recent generalizations of the results by Liu and Wang in [68], −
u + V (x)u = f (x, u). Throughout this section we assume (V3 ), (f0′ ), (f2 ) and

(3.7)

46

T. Bartsch et al.

(V′0 ) V ∈ C 1 (RN , R) and infRN V (x) > 0. (f6 ) f ∈ C 1 (RN × R, R) is 1-periodic in each of its x-variables. The periodicity conditions imply that (3.7) is ZN -invariant. The weak solutions of (3.7) correspond to critical points of  

 1 Φ(u) := |∇u|2 + V (x)u2 dx − F (x, u) dx 2 RN RN

in E = H 1 (RN ). Define the mountain pass value c as

 c = inf sup Φ g(t) , g∈Γ t∈[0,1]

where

 

 Γ = g ∈ C [0, 1], E  g(0) = 0, Φ g(1) < 0 .

In the homogeneous case f (x, u) = |u|p−2 u this is just the minimizer of Φ on the Nehari manifold. In [48] it was proved that (3.7) has infinitely many k-bump solutions. In particular, we have the following theorem. kc+α T HEOREM 3.21. Kkc−α /ZN is infinite for any k
2, provided the following condition is satisfied: (∗) there is α > 0 such that Kc+α /ZN is finite.

Here Kab denotes the set of critical points of Φ between the levels a and b. kc+α In [48,68] it is proved that Kkc−α /ZN contains infinitely many nodal solutions. There one can also find estimates on the number of nodal domains of these multibump type solutions. In the following we review the modified approach given by Liu and Wang in [68]. This modified approach gives the one sign solutions and the nodal solutions in a uniform fashion, and it is easier for obtaining estimates on the number of nodal domains compared with the original approach by Coti Zelati and Rabinowitz [48]. For a > 0, the a-neighborhood of a set A ⊂ E is defined by  Na (A) = u ∈ E: u − A < a ,

a (A) and ∂Na (A), respectively. For j = whose closure and boundary are denoted by N (j1 , . . . , jN ) ∈ ZN , we define the translation on RN by τj u(x) = u(x1 + j1 , . . . , xN + jN ). For a finite subset E1 of E and an integer l
1, we denote Tl (E1 ) =



j  i=1



τki vi : 1  j  l, vi ∈ E1 , ki ∈ ZN .

The Dirichlet problem for superlinear elliptic equations

47

Consider the positive cone P + and the negative cone P − in E defined by P ± = {u ∈ E: ±u
0}. Any u ∈ K \ (P + ∪ P − ) will be a nodal solution of (3.7). L EMMA 3.22. (i) there is ν > 0 such that u
ν for all u ∈ K \ {0}; (ii) there is c > 0 such that Φ(u)
c for all u ∈ K \ {0}; (iii) for all u ∈ K \ {0} with Φ(u)  b: u  (2θ b/(θ − 2))1/2 ; (iv) for any b > 0, there is ν1 > 0 depending on b such that u± L2 (RN )
ν1 for all u ∈ K \ (P + ∪ P − ) with Φ(u)  b. See the proofs in [48] and [68]. Let A : E → E be given by A(u) := (−
+ V )−1 [f (·, u(·))] for u ∈ E. Then the gradient of Φ has the form Φ ′ (u) = u − A(u). Note that the set of fixed points of A is the same as the set of critical points of Φ, which is K. By the proof of [48], Proposition 2.1, Φ ′ : E → E is locally Lipschitz continuous. Using Lemma 3.17, the behavior of PS sequences in the whole space E as well as a (P ± ) can be studied. in N L EMMA 3.23. Let (um ) ⊂ E be such that Φ(um ) → b > 0 and Φ ′ (um ) → 0. Then there is an l ∈ N (depending on b), v1 , . . . , vl ∈ K \ {0}, a subsequence of um and corresponding i ∈ ZN , i = 1, . . . , l, such that km   l      τkmi vi  → 0 um −   i=1

l  i=1

as m → ∞,

Φ(vi ) = b,

(3.8)

(3.9)

and for i = j ,

 i j  k − k m → ∞ as m → ∞. m

(3.10)

a1 (P + ) Moreover, there exists an a1 ∈ (0, a0 ] (depending on b) such that if (um ) ⊂ N − + − (Na1 (P ), resp.) then v1 , . . . , vl ∈ (K \ {0}) ∩ P ((K \ {0}) ∩ P , resp.). For a ∈ [0, a1 ], we define

and





 a P ± : g(0) = 0 and Φ g(1) < 0 Γa± = g ∈ C [0, 1], N

 ca± = inf max Φ g(θ ) . g∈Γa± θ∈[0,1]

48

T. Bartsch et al.

a (P ± ) = P ± . In this case, we denote Γ ± = Γ ± and c± = c± . Then we can For a = 0, N 0 0 ± prove that ca is independent of a for a small. L EMMA 3.24. There exists a2 ∈ (0, a1 ) such that ca± = c± for all a ∈ (0, a2 ]. Denote Ki = K ∩ P i for i ∈ {+, −}. We will also use the notation: (Ki )b = Ki ∩ Φ b , (Ki )ba = Ki ∩ Φab and Ki (ci ) = K(ci ) ∩ P i for i ∈ {+, −}. Instead of (∗), we need the following condition. ± (∗)± There is α > 0 such that (K± )c +α /ZN is finite. i Choose a representative in E from each equivalent class in (Ki )c +α /ZN and denote the resulting set by F i , i ∈ {+, −}. Let c > 0 be the number from Lemma 3.22 which satisfies Φ(u)
c for all u ∈ K \ {0}. Denote l ± = [(c± + α)/c]. According to [48], Proposition 2.57, µ(Tl ± (F ± )) = inf{u − w: u = w ∈ Tl ± (F ± )} > 0. Using this a deformation a (P ± ) can be given. The proof of the deformation lemma is similar to the lemma in N a (P i ) proof of [48]. However, we need to construct a descending flow of Φ which makes N i a (P ) to itself. positively invariant so that the deformation is from N Then by using the descending flow, one obtains the following theorem which asserts the existence of one-bump positive and negative solutions at the mountain pass level. These one-bump solutions will be used later to construct multibump nodal solutions. T HEOREM 3.25. Let (∗)± be satisfied. Then c± are critical values of Φ and there are critical points u± ∈ K± such that Φ(u± ) = c± . R EMARK 3.26. This theorem shows the existence of a solution near the mountain pass level regardless of whether (∗)± is satisfied or not. Now, by (∗)± , there is an α1 ∈ (0, α) such that

i ci +α1

 K ci −α = Ki ci . 1

L EMMA 3.27. Let (∗)± be satisfied. Then there exist finite sets A+ ⊂ K+ (c+ ) and 1 µ(Tl ± (F ± )) and p ∈ N, A− ⊂ K− (c− ) having the property that for any ε¯ 1  α1 /2, r1  12 ± ± there is an ε1 ∈ (0, ε¯ 1 ) and g1 ∈ Γ such that (i) maxθ∈[0,1] Φ(g1± (θ ))  c± + ε1 /p; (ii) if Φ(g1± (θ )) > c± − ε1 then g1± (θ ) ∈ Nr1 (A± ). Let A = A+ ∪ A− with A± given in Lemma 3.27. For any fixed integer k
2 we fix two positive integers k + and k − such that k = k + + k − . Denote Λ+ = {1, . . . , k + }, Λ− = {k + + 1, . . . , k}. Let ji ∈ ZN for i = 1, . . . , k be fixed such that ji = jm for i = m and if vi ∈ A+ for i ∈ Λ+ and vi ∈ A− for i ∈ Λ− then  k    kν   τ ji v i 
   2 i=1

  k     α  

+ +   and Φ τ ji v i − k c + k − c −  < .   2 i=1

The Dirichlet problem for superlinear elliptic equations

49

Define

 M j1 , . . . , jk , A, k + , k −   k  + + − − τji vi : vi ∈ A for i ∈ Λ , vi ∈ A for i ∈ Λ = i=1

and bk = k + c + + k − c − . The main theorem in [68] reads as follows. T HEOREM 3.28. Let (∗)± be satisfied. Then there is an r0 > 0 such that, for any r ∈ (0, r0 ),

 b +α   Nr M lj1 , . . . , ljk , A, k + , k − ∩ Kbkk −α ZN = ∅

for all but finitely many l ∈ N.

We sketch the proof of Theorem 3.28. For θ = (θ1 , . . . , θk ) ∈ [0, 1]k , let 0i = (θ1 , . . . , θi−1 , 0, θi+1 , . . . , θk ) and 1i = (θ1 , . . . , θi−1 , 1, θi+1 , . . . , θk ), 1  i  k. Let a2 be as in Lemma 3.24 and a ∈ [0, a2 ] and define  Γk (a) = G = g1 + · · · + gk : gi satisfies (g1 )–(g3 ), 1  i  k ,

where a (P ± )) for i ∈ Λ± , (g1 ) gi ∈ C([0, 1]k , N (g2 ) gi (0i ) = 0 and Φ(gi (1i )) < 0, 1  i  k, i ∩ O j = ∅ if i = j and (g3 ) there are bounded open sets Oi , 1  i  k, such that O k supp gi (θ ) ⊂ Oi for all θ ∈ [0, 1] . L EMMA 3.29. Let (∗)± be satisfied. Define

 bk (a) = inf max Φ G(θ ) . G∈Γk (a) θ∈[0,1]k

Then bk (a) = bk = k + c+ + k − c− for a ∈ (0, a2 ]. From here on we argue by contradiction. Thus assume that there are infinitely many l such that the conclusion of the theorem is false. The proofs go with the construction of a special G, which is chosen as close to the level bk (a) as possible, the deformation of this special G under the gradient flow, modification of the deformed G so that finally we get a map in the family Γk (a) satisfying max Φ(G) < bk , which is a contradiction with the definition of bk (a). We refer for the details to [68]. Next we state a result of [48,68] which is used in establishing more detailed nodal properties of these solutions.

50

T. Bartsch et al.

T HEOREM 3.30. In the above theorem, if r is sufficiently small and l sufficiently large, we may replace the r-neighborhood in E with r-neighborhood in X := C 1 (RN ). The proof of this uses some elliptic estimates. T HEOREM 3.31. Suppose (∗)± holds. For multibump nodal solutions of (3.7), the number of nodal domains is bounded by the number of bumps. In particular, the two-bump nodal solutions have exactly two nodal domains. Moreover, there are infinitely many, geometrically different, two-bump, nodal solutions which have exactly two nodal domains. The proof of this theorem is a consequence of the following lemma from [68]. L EMMA 3.32. If r is small enough and l is large enough then, for the solutions u given in Theorem 3.28, the number of nodal domains is bounded above by the number of bumps of the solutions. The techniques for the above theorem can be used to establish the nodal property of the solutions in another two cases. First, we consider (3.7) on a cylinder domain Ω = ω × R and we write x = (x ′ , xN ) with x ′ = (x1 , . . . , xN −1 ), where ω is a bounded smooth domain in RN −1 . We assume (V0 ), (f0′ ), (f2 ), (f3 ), and (V5 ) and (f6 ) are satisfied with the periodicity only in the xN direction. The space will be E = H01 (Ω). Then we can still define the mountain pass values c± > 0. The problem now is Z invariant. ± (∗′ )± There is α > 0 such that Kc +α /Z is finite. T HEOREM 3.33. Suppose (∗′ )± holds. Then, for any integers k
m
2, (3.7) has inkc+α finitely many, geometrically different, k-bump, nodal solutions in Ikc−α which have exactly m nodal domains. More precisely, given any positive integers k , k , . . . , km such that 1 2 m i=1 ki = k
2, there are infinitely many, geometrically different, k-bump, nodal solukc+α tions in Ikc−α which have exactly m nodal domains Di , i = 1, . . . , m, such that u|Di is a ki -bump positive or negative solution. Another case can be considered is when (3.7) has different x-dependence in different directions. We assume (V0 ), (f0′ ), (f2 ), (f3 ) and (V5 ) V is periodic in xN and radially symmetric in (x1 , . . . , xN −1 ). (f7 ) f ∈ C 1 (RN × R, R) is periodic in xN and radially symmetric in (x1 , . . . , xN −1 ). Then Z acts via translations on the xN -variable, and the problem is Z-invariant. With x = (x ′ , xN ) and x ′ = (x1 , . . . , xN −1 ), we work with the space
  

 ′ 

N ′ 1   E = u ∈ H R : u x , xN = u x , xN ,

RN

V (x)u dx < ∞ , 2

that is, functions in E are radially symmetric in the first N − 1 variables. We can still define the mountain pass values c± in the space E. ± (∗′′ )± There is α > 0 such that Kc +α /Z is finite.

The Dirichlet problem for superlinear elliptic equations

51

T HEOREM 3.34. Suppose (∗′′ )± holds. Then for any integer k
2, (3.7) has infinitely kc+α many, geometrically different, k-bump nodal solutions in Φkc−α such that the numbers of their nodal domains are bounded between [k/2] + 1 and k. In particular, there are nodal solutions such that the numbers of their nodal domains tend to infinity. We refer the details of the proofs to [68].

3.6. Strongly indefinite potentials In this section we consider problem (3.1) when V and f are periodic in the x-variables but V is not positive anymore. For periodic V the operator −
+ V has a band spectrum. More precisely, the spectrum of −
+ V on L2 (RN ) is purely continuous, bounded below and consists of a disjoint union of closed intervals. We require: (V5 ) V ∈ C(RN , R) is 1-periodic in each of its variables and the linear operator, H 2 (RN ) → L2 (RN ),

u → −
u + V (x)u,

is invertible. By (V5 ) the operator −
+ V (x) is self-adjoint on L2 (RN ) and may have continuous spectrum on the negative real axis with 0 in a spectral gap. T HEOREM 3.35. Suppose (V5 ) and (f0 ) with p < 2∗ , and (f2 ) hold. Then problem (3.1) has a nontrivial solution in H 1 (RN ). Theorem 3.35 is due to Kryszewski and Szulkin [58]. An earlier version requiring a stronger growth condition on f has been obtained by Troestler and Willem [89]. If 0 is a left end point of a spectral gap, that is 0 ∈ σ (−
+ V ) and (0, ε) ⊂ R \ σ (−
+ V ) then a solution of (3.1) has been found in [11] under additional growth conditions on f . The 2 (RN ) ∩ Lq (RN ) for p  q  2∗ . solution then does not lie in H 1 (RN ) but only in Hloc T HEOREM 3.36. Suppose (V5 ) and (f0 ) with p < 2∗ , and (f2 ) hold. If in addition f is odd in u then problem (3.1) has infinitely many geometrically different solutions in H 1 (RN ). Here two solutions u1 , u2 are said to be geometrically different if there does not exist k ∈ Zn with u2 (x) = u1 (x + k) for every x ∈ RN . Also Theorem 3.36 is due to [58] and has been extended in [11] to the case where 0 is a left end point of a spectral gap.

Acknowledgments The project was supported by NATO grant PST.CLG.978763. The second author would like to thank the hospitality of the Institute for Mathematics and its Applications at the University of Minnesota and the Institute of Mathematical Sciences at the Chinese University of Hongkong where the paper was finished during his sabbatical visits.

52

T. Bartsch et al.

References [1] A. Aftalion and F. Pacella, Qualitative properties of nodal solutions of semilinear elliptic equations in radially symmetric domains, C. R. Math. Acad. Sci. Paris 339 (2004), 339–344. [2] A. Ambrosetti, Critical points and nonlinear variational problems, Mém. Soc. Math. France Sér. 2 49 (1992), 1–139. [3] A. Ambrosetti, H. Brezis and G. Cerami, Combined effects of concave and convex nonlinearities in some elliptic problems, J. Funct. Anal. 122 (1994), 519–543. [4] A. Ambrosetti, J. Garcia Azorero and I. Peral, Multiplicity results for some nonlinear elliptic equations, J. Funct. Anal. 137 (1996), 219–242. [5] A. Ambrosetti and P.H. Rabinowitz, Dual variational methods in critical point theory and applications, J. Funct. Anal. 14 (1973), 349–381. [6] A. Bahri, Topological results on a certain class of functionals and applications, J. Funct. Anal. 14 (1981), 397–427. [7] A. Bahri and H. Berestycki, A perturbation method in critical point theory and applications, Trans. Amer. Math. Soc. 267 (1981), 1–32. [8] A. Bahri and J.-M. Coron, On a nonlinear elliptic equation involving the critical Sobolev exponent, Comm. Pure Appl. Math. 41 (1988), 255–294. [9] A. Bahri and P.L. Lions, Morse index of some min–max critical points, I. Applications to multiplicity results, Comm. Pure Appl. Math. 41 (1988), 1027–1037. [10] T. Bartsch, Critical point theory on partially ordered Hilbert spaces, J. Funct. Anal. 186 (2001), 117–152. [11] T. Bartsch and Y. Ding, On a nonlinear Schrödinger equation with periodic potential, Math. Ann. 313 (1999), 15–37. [12] T. Bartsch, Z. Liu and T. Weth, Sign changing solutions to superlinear Schrödinger equations, Comm. Partial Differential Equations 29 (2004), 25–42. [13] T. Bartsch, A. Pankov and Z.-Q. Wang, Nonlinear Schrödinger equations with steep potential well, Commun. Contemp. Math. 3 (2001), 1–21. [14] T. Bartsch and Z.-Q. Wang, Existence and multiplicity results for some superlinear elliptic problems on RN , Comm. Partial Differential Equations 20 (1995), 1725–1741. [15] T. Bartsch and Z.-Q. Wang, On the existence of sign changing solutions for semilinear Dirichlet problems, Topol. Methods Nonlinear Anal. 7 (1996), 115–131. [16] T. Bartsch and Z.-Q. Wang, Multiple positive solutions for a nonlinear Schrödinger equation, Z. Angew. Math. Phys. 51 (2000), 366–384. [17] T. Bartsch and T. Weth, A note on additional properties of sign changing solutions to superlinear Schrödinger equations, Topol. Methods Nonlinear Anal. 22 (2003), 1–14. [18] T. Bartsch and T. Weth, Three nodal solutions of singularly perturbed elliptic equations on domains without topology, Ann. Inst. H. Poincaré Anal. Non Linéaire, to appear. [19] T. Bartsch, T. Weth and M. Willem, Partial symmetry of least energy nodal solutions to some variational problems, J. Anal. Math., to appear. [20] T. Bartsch and M. Willem, Infinitely many nonradial solutions of a Euclidean scalar field equation, J. Funct. Anal. 117 (1993), 447–460. [21] V. Benci and G. Cerami, The effect of the domain topology on the number of positive solutions of nonlinear elliptic problems, Arch. Ration. Mech. Anal. 114 (1991), 79–93. [22] V. Benci and G. Cerami, Multiple positive solutions of some elliptic problems via the Morse theory and the domain topology, Calc. Var. Partial Differential Equations 2 (1994), 29–48. [23] V. Benci, G. Cerami and D. Passaseo, On the number of positive solutions of some nonlinear elliptic problems, Nonlinear Analysis a Tribute in Honour of G. Prodi, A. Ambrosetti and A. Marino, eds, Quaderni Scuola Normale, Pisa (1993), 93–107. [24] H. Berestycki, Qualitative properties of positive solutions of elliptic equations, Partial Differential Equations, Prague (1998), Chapman & Hall/CRC Res. Notes Math., Vol. 406, Chapman & Hall/CRC, Boca Raton, FL (2000), 34–44. [25] H. Berestycki and P.-L. Lions, Nonlinear scalar field equations, I. Existence of a ground state, Arch. Ration. Mech. Anal. 82 (1983), 313–345.

The Dirichlet problem for superlinear elliptic equations

53

[26] H. Berestycki and L. Nirenberg, On the method of moving planes and the sliding method, Bol. Soc. Brasil Mat. (N.S.) 22 (1991), 1–37. [27] Ph. Bolle, On the Bolza problem, J. Differential Equations 152 (1999), 274–288. [28] Ph. Bolle, N. Ghoussoub and H. Tehrani, The multiplicity of solutions to non-homogeneous boundary value problems, Manuscripta Math. 101 (2000), 325–350. [29] H. Brezis, Symmetry in nonlinear PDE’s, Differential Equations, La Pietra, 1996 (Florence), Proc. Sympos. Pure Math., Vol. 65, Amer. Math. Soc., Providence, RI (1999), 1–12. [30] H. Brezis and T. Kato, Remarks on the Schrödinger operator with singular complex potentials, J. Math. Pures Appl. 58 (1979), 137–151. [31] H. Brezis and E. Lieb, A relation between pointwise convergence of functions and convergence of functionals, Proc. Amer. Math. Soc. 88 (1983), 486–490. [32] H. Brezis and L. Nirenberg, Positive solutions of nonlinear elliptic equations involving critical Sobolev exponents, Comm. Pure Appl. Math. 36 (1983), 437–477. [33] H. Brezis and R.E.L. Turner, On a class of superlinear elliptic problems, Comm. Partial Differential Equations 2 (1977), 601–614. [34] F. Brock and P. Solynin, An approach to symmetrization via polarization, Trans. Amer. Math. Soc. 352 (2000), 1759–1796. [35] J. Byeon, Existence of many nonequivalent nonradial positive solutions of semilinear elliptic equations on three-dimensional annuli, J. Differential Equations 136 (1997), 136–165. [36] A. Castro, J. Cossio and J.M. Neuberger, A sign-changing solution for a superlinear Dirichlet problem, Rocky Mountain J. Math. 27 (1997), 1041–1053. [37] A. Castro, J. Cossio and J.M. Neuberger, A minmax principle, index of the critical point, and existence of sign-changing solutions to elliptic BVPs, Electron. J. Differential Equations 2 (1998), 1–18. [38] F. Catrina and Z.-Q. Wang, Nonlinear elliptic equations on expanding symmetric domains, J. Differential Equations 156 (1999), 153–181. [39] J. Chabrowski, Variational Methods for Potential Operator Equations, de Gruyter, Berlin (1997). [40] J. Chabrowski, Weak Convergence Methods for Semilinear Equations, World Scientific, River Edge, NJ (1999). [41] K.C. Chang, Infinite Dimensional Morse Theory and Multiple Solution Problems, Birkhäuser, Boston (1993). [42] C.V. Coffman, Uniqueness of the ground state solution for
u−u+u3 = 0 and variational characterization of other solutions, Arch. Ration. Mech. Anal. 49 (1972), 81–95. [43] C.V. Coffman, A nonlinear boundary value problem with many positive solutions, J. Differential Equations 54 (1984), 429–437. [44] M. Conti, L. Merizzi and S. Terracini, Remarks on variational methods and lower–upper solutions, NoDEA Nonlinear Differential Equations Appl. 6 (1999), 371–393. [45] M. Conti, S. Terracini and G. Verzini, Nehari’s problem and competing species systems, Ann. Inst. H. Poincaré Anal. Non Linéaire 19 (2002), 871–888. [46] D.G. Costa and C.A. Magalhaes, Variational elliptic problems which are nonquadratic at infinity, Nonlinear Anal. 23 (1994), 1401–1412. [47] D.G. Costa and Z.-Q. Wang, Multiplicity results for a class of superlinear elliptic problems, Proc. Amer. Math. Soc. 133 (2005), 787–794. [48] V. Coti Zelati and P. Rabinowitz, Homoclinic type solutions for a semilinear elliptic PDE on Rn , Comm. Pure Appl. Math. 45 (1992), 1217–1269. [49] E. Dancer and Y. Du, Existence of sign-changing solutions for some semilinear problems with jumping nonlinearities, Proc. Roy. Soc. Edinburgh Sect. A 124 (1994), 1165–1176. [50] E. Dancer and Y. Du, On sign changing solutions of certain semi-linear elliptic problems, Appl. Anal. 56 (1995), 193–206. [51] D.G. de Figueiredo, P.L. Lions and R. Nussbaum, A priori estimates and existence of positive solutions of semilinear elliptic equations, J. Math. Pures Appl. 61 (1982), 41–63. [52] N. Ghoussoub, Duality and Perturbation Methods in Critical Point Theory, Cambridge University Press, Cambridge (1993). [53] B. Gidas, W.-M. Ni and L. Nirenberg, Symmetry and related properties via the maximum principle, Comm. Math. Phys. 68 (1979), 209–243.

54

T. Bartsch et al.

[54] B. Gidas and J. Spruck, A priori bounds for positive solutions of nonlinear elliptic equations, Comm. Partial Differential Equations 6 (1981), 883–901. [55] L. Jeanjean and K. Tanaka, A remark on least energy solutions in RN , Proc. Amer. Math. Soc. 131 (2003), 2399–2408. [56] O. Kavian, Introduction à la théorie des points critiques et applications aux problèmes elliptiques, SpringerVerlag, Heidelberg (1993). [57] H. Kielhöfer, Bifurcation Theory—An Introduction with Applications to PDEs, Springer-Verlag, New York (2004). [58] W. Kryzewski and A. Szulkin, Generalized linking theorem with an application to semilinear Schrödinger equation, Adv. Differential Equations 3 (1998), 441–472. [59] M.R. Kwong, Uniqueness of positive solutions of
u − u + up = 0 in Rn , Arch. Ration. Mech. Anal. 105 (1989), 243–266. [60] S.J. Li and Z.-Q. Wang, Mountain pass theorems in order intervals and multiple solutions for semilinear elliptic Dirichlet problems, J. Anal. Math. 81 (2001), 373–396. [61] S.J. Li and Z.-Q. Wang, Lusternik–Schnirelman theory in partially ordered Hilbert spaces, Trans. Amer. Math. Soc. 354 (2002), 3207–3227. [62] Y.Y. Li, Existence of many positive solutions of semilinear elliptic equations on annulus, J. Differential Equations 83 (1990), 348–367. [63] Y. Li and Z. Liu, Multiple and sign changing solutions of an elliptic eigenvalue problem with constraint, Sci. China Ser. A 44 (2001), 48–57. [64] P. Lions, The concentration–compactness principle in the calculus of variations the locally compact case. Part II, Ann. Inst. H. Poincaré Anal. Non Linéaire 1 (1984), 223–283. [65] Z. Liu and J. Sun, Invariant sets of descending flow in critical point theory with applications to nonlinear differential equations, J. Differential Equations 172 (2001), 257–299. [66] Z.-L. Liu, F.A. van Heerden and Z.-Q. Wang, Nodal type bound states of Schrödinger equations via invariant set and minimax methods, J. Differential Equations 214 (2005), 358–390. [67] Z.-L. Liu and Z.-Q. Wang, On the Ambrosetti–Rabinowitz superlinear condition, Adv. Nonlinear Stud. 4 (2004), 561–572. [68] Z.-L. Liu and Z.-Q. Wang, Multi-bump, nodal solutions having a prescribed number of nodal domains, I and II, Ann. Inst. H. Poincaré, Anal. Non Linéaire, to appear. [69] Z.-L. Liu and Z.-Q. Wang, On an elliptic equation on RN with concave and convex nonlinearities, Z. Angew. Math. Phys., to appear. [70] O. Lopes, Radial and nonradial minimizers for some radially symmetric functionals, Electron. J. Differential Equations 1966 (1966), 1–14. [71] S. Lorca and P. Ubilla, Symmetric and nonsymmetric solutions for an elliptic equation on RN , Nonlinear Anal. 58 (2004), 961–968. [72] E. Montefusco, Axial symmetry of solutions to semilinear elliptic equations in unbounded domains, Proc. Roy. Soc. Edinburgh Sect. A 133 (2003), 1175–1192. [73] E. Müller-Pfeiffer, On the number of nodal domains for elliptic differential operators, J. London Math. Soc. 31 (1985), 91–100. [74] Z. Nehari, On a nonlinear differential equation arising in nuclear physics, Proc. Roy. Irish Acad. Sect. A 62 (1963), 117–135. [75] W.-M. Ni and J.C. Wei, On the location and profile of spike-layer solutions to singularly perturbed semilinear Dirichlet problems, Comm. Pure Appl. Math. 48 (1995), 731–768. [76] F. Pacella, Symmetry results for solutions of semilinear elliptic equations with convex nonlinearities, J. Funct. Anal. 192 (2002), 271–282. [77] R.S. Palais, The principle of symmetric criticality, Comm. Math. Phys. 69 (1979), 19–30. [78] S. Pohozaev, Eigenfunctions of the equation
u + λf (u) = 0, Soviet. Math. Dokl. 6 (1965), 1408–1411. [79] P.H. Rabinowitz, Minimax Methods in Critical Point Theory with Applications to Differential Equations, CBMS Reg. Conf. Ser. Math., Vol. 65, Amer. Math. Soc., Providence, RI (1986). [80] M. Ramos, Z.-Q. Wang and M. Willem, Positive solutions for elliptic equations with critical growth in unbounded domains, Calculus of Variations and Differential Equations, Haifa, 1998, Chapman & Hall/CRC Res. Notes Math., Vol. 410, Chapman &Hall/CRC, Boca Raton, FL (2000), 192–199. [81] M. Schechter, Linking Methods in Critical Point Theory, Birkhäuser, Boston (1999).

The Dirichlet problem for superlinear elliptic equations

55

[82] M. Schechter and W. Zou, Superlinear problems, Pacific J. Math. 214 (2004), 145–160. [83] D. Smets and M. Willem, Partial symmetry and asymptotic behaviour for some elliptic variational problems, Calc. Var. Partial Differential Equations 18 (2003), 57–75. [84] P.N. Srikanth, Uniqueness of solutions of nonlinear Dirichlet problems, Differential Integral Equations 6 (1993), 663–670. [85] W.A. Strauss, Existence of solitary waves in higher dimensions, Comm. Math. Phys. 55 (1977), 149–162. [86] M. Struwe, Infinitely many critical points for functionals which are not even and applications to nonlinear boundary value problems, Manuscripta Math. 32 (1980), 335–364. [87] M. Struwe, Multiple solutions of anticoercive boundary value problems for a class of ordinary differential equations of second order, J. Differential Equations 37 (1980), 285–295. [88] M. Struwe, Variational Methods, Springer-Verlag, Berlin (1990). [89] C. Troestler and M. Willem, Nontrivial solution of a semilinear Schrödinger equation, Comm. Partial Differential Equations 21 (1996), 1431–1449. [90] Z.-Q. Wang, On a superlinear elliptic equation, Ann. Inst. H. Poincaré Anal. Non Linéaire 8 (1991), 43–57. [91] Z.-Q. Wang, Existence and symmetry of multi-bump solutions for nonlinear Schrödinger equations, J. Differential Equations 159 (1999), 102–137. [92] Z.-Q. Wang, Minimax methods, invariant sets, and applications to nodal solutions of nonlinear elliptic problems, Proceedings of EquaDiff 2003, Hasselt, Belgium, 2003, World Scientific, Singapore (2005), 561–566. [93] Z.-Q. Wang and M. Willem, Existence of many positive solutions of an elliptic equation with critical exponent on an annulus, Proc. Amer. Math. Soc. 127 (1999), 1711–1714. [94] J. Wei and M. Winter, Symmetry of nodal solutions for singularly perturbed elliptic problems on a ball, Preprint (2004). [95] M. Willem, Minimax Theorems, Birkhäuser, Boston (1996). [96] M. Willem, Analyse fonctionnelle élémentaire, Cassini, Paris (2003).

This page intentionally left blank

CHAPTER 2

Nonconvex Problems of the Calculus of Variations and Differential Inclusions Bernard Dacorogna Section of Mathematics, EPFL, 1015 Lausanne, Switzerland E-mail: [email protected]

Contents 1. Introduction . . . . . . . . . . . . . . . . . . . 2. Preliminaries and notations . . . . . . . . . . 2.1. The different notions of convexity . . . . 2.2. Some function spaces . . . . . . . . . . . 2.3. Statement of the problem . . . . . . . . . 3. Relaxation theorems . . . . . . . . . . . . . . 3.1. The different envelopes . . . . . . . . . . 3.2. Some examples . . . . . . . . . . . . . . 3.3. The main theorem . . . . . . . . . . . . . 4. Implicit partial differential equations . . . . . 4.1. Introduction . . . . . . . . . . . . . . . . 4.2. The different convex hulls . . . . . . . . 4.3. Some examples of convex hulls . . . . . 4.4. An existence theorem . . . . . . . . . . . 4.5. Some examples of existence of solutions 4.6. Appendix . . . . . . . . . . . . . . . . . 5. Existence of minimizers . . . . . . . . . . . . 5.1. Introduction . . . . . . . . . . . . . . . . 5.2. Sufficient conditions . . . . . . . . . . . 5.3. Necessary conditions . . . . . . . . . . . 6. The scalar case . . . . . . . . . . . . . . . . . 6.1. The case of single integrals . . . . . . . . 6.2. The case of multiple integrals . . . . . . 7. The vectorial case . . . . . . . . . . . . . . . . 7.1. The case of singular values . . . . . . . . 7.2. The case of quasiaffine functions . . . . . 7.3. The Saint Venant–Kirchhoff energy . . . 7.4. An optimal design problem . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

HANDBOOK OF DIFFERENTIAL EQUATIONS Stationary Partial Differential Equations, volume 2 Edited by M. Chipot and P. Quittner © 2005 Elsevier B.V. All rights reserved 57

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

59 62 62 66 66 66 67 68 70 74 74 74 75 77 81 83 86 86 86 92 104 105 107 110 111 113 115 116

58

B. Dacorogna

7.5. The minimal surface case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 7.6. The problem of potential wells . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

Abstract We study existence of minimizers for problems of the type


 

 1,p N , f x, u(x), Du(x) dx: u ∈ u0 + W0 Ω; R inf

(P)

Ω

where u0 is a given function. After recalling some basic facts about existence of minimizers when the function f is convex (quasiconvex), we turn our attention to the case where f is not convex (quasiconvex). We start by presenting the general tool of relaxation, which gives generalized solutions of (P). We next discuss some differential inclusions, where we look for solutions u ∈ u0 + W01,∞ (Ω; RN ) of Du(x) ∈ E

a.e. in Ω,

where E ⊂ RN ×n is a given compact set. Finally combining the relaxation theorem and the study of differential inclusions, we give necessary and sufficient conditions for existence of classical minimizers of (P) as well as several examples.

Nonconvex problems of the calculus of variations and differential inclusions

59

1. Introduction We discuss the existence of minimizers for the problem




  1,p inf I (u) = f x, u(x), Du(x) dx: u ∈ u0 + W0 Ω; RN ,

(P)

Ω

where • Ω ⊂ Rn is a bounded open set, with Lipschitz boundary ∂Ω; • u : Ω → RN ,

 u = u(x) = u(x1 , . . . , xn ) = u1 (x), . . . , uN (x)

(if N = 1 or, by abuse of language, if n = 1, we will say that it is scalar valued while if N, n
2, we will speak of the vector valued case); • Du denotes its Jacobian matrix, i.e., Du =



∂ui ∂xj

1iN 1j n



∂u1 ∂x1

··· .. .

∂u1 ∂xn

∂uN ∂x1

···

∂uN ∂xn

 =  ...



..  ∈ RN ×n ; . 

• f : Ω × RN × RN ×n → R is continuous, f = f (x, u, ξ ); • 1  p  ∞ and W 1,p (Ω; RN ) denotes the usual space of Sobolev maps, where ui ,

∂ui ∈ Lp (Ω), ∂xj

i = 1, . . . , N, j = 1, . . . , n;

• u0 ∈ W 1,p (Ω; RN ) is a given map; 1,p • u ∈ u0 + W0 (Ω; RN ), meaning that u ∈ W 1,p (Ω; RN ) and u = u0 on ∂Ω in the Sobolev sense. This problem is the fundamental problem of the calculus of variations and it has received a considerable attention since the time of Fermat, Newton, Bernoulli, Euler and all along the 19th and 20th centuries. The most general way of proving existence of minimizers of (P), meaning to find u¯ ∈ 1,p u0 + W0 (Ω; RN ) so that I (u) ¯  I (u) 1,p

among all admissible u ∈ u0 + W0 (Ω; RN ), is the so-called direct methods of the calculus of variations. These methods rely on some kind of convexity condition of the function ξ → f (x, u, ξ ). There are numerous examples showing that in absence of convexity the problem (P) has no minimizers. At the moment let us quote three elementary examples where nonexistence occurs.

60

B. Dacorogna

E XAMPLE 1. Let N = n = 1, f (x, u, ξ ) = f (ξ ) = e−ξ

2

and
 inf I (u) =

1

0



′  1,1 f u (x) dx: u ∈ W0 (0, 1) .

(P)

E XAMPLE 2. Let N = n = 1,

and

2 f (x, u, ξ ) = f (u, ξ ) = u4 + ξ 2 − 1
 inf I (u) =

1

0



 f u(x), u′ (x) dx: u ∈ W01,4 (0, 1) .

(P)

This example is due to Bolza. E XAMPLE 3. Let n = 2, N = 1, Ω = (0, 1)2 ,

and

2 f (x, u, ξ ) = f (ξ ) = f (ξ1 , ξ2 ) = ξ12 − 1 + ξ24




 1,4 inf I (u) = f Du(x) dx: u ∈ W0 (Ω) .

(P)

Ω

We now continue this introduction by discussing only the scalar case (i.e., when N = 1 or n = 1), the general vectorial case will be discussed in the next sections. We moreover, in order to simplify the presentation, consider the case where there is no dependence on lower-order terms, i.e., f (x, u, ξ ) = f (ξ ). When dealing with nonconvex problems, the first step is the relaxation theorem, established by L.C. Young, Mac Shane, Ekeland and others. This consists in replacing the problem (P) by the so-called relaxed problem




 1,p  inf I (u) = Cf Du(x) dx: u ∈ u0 + W0 (Ω) , Ω

where Cf is the convex envelope of f , namely Cf = sup{g  f : g convex}.

(QP)

Nonconvex problems of the calculus of variations and differential inclusions

61

Therefore the direct methods, which do not apply to (P), apply to (QP). It can be shown (cf. Theorem 15) that inf(P) = inf(QP) and that minimizers of (P) are necessarily minimizers of (QP), the converse being false. In the three above examples we have (i) Cf (ξ ) ≡ 0, inf(P) = inf(QP) = 0 and any u ∈ W01,1 (0, 1) is a solution of (QP); (ii) Cf (u, ξ ) = u4 + [ξ 2 − 1]2+ , inf(P) = inf(QP) = 0 and u ≡ 0 is a solution of (QP); (iii) Cf (ξ ) = [ξ12 − 1]2+ + ξ24 , inf(P) = inf(QP) = 0 and u ≡ 0 is a solution of (QP); where, for x ∈ R,
x if x
0, [x]+ = 0 if x < 0. The second step in proving the existence of minimizers for (P) is to see if among all solutions of (QP), if any, at least one of them is also a solution of (P). This amounts in 1,p finding u¯ ∈ u0 + W0 (Ω) so that 

Ω

 Cf D u(x) ¯ dx = inf(P) = inf(QP)

and at the same time in solving the first-order differential equation (called, following Dacorogna and Marcellini [31], implicit partial differential equation)



 Cf D u(x) ¯ = f D u(x) ¯

a.e. x ∈ Ω.

After this brief and informal introduction, we discuss the organization of the article. In Section 2 we discuss all the notions of convexity that are involved in the vector valued case, in particular the so-called quasiconvexity. In Section 3 we present the relaxation theorem in the vector valued case, introducing all the needed generalization of the notion of convex envelope. In Section 4 we give some existence theorems for implicit differential equations of the above type. In Section 5 using the results of the two preceding sections, we discuss necessary and sufficient conditions for the existence of minimizers for nonconvex problems. In Section 6 we show how to apply the abstract results to scalar problems; obtaining sharper theorems in the case of single integrals (i.e., n = 1). In Section 7 we present several examples involving vector valued functions (i.e., n, N
2) which are relevant for applications. The subject is very large and we do not intend to be complete and we refer to the bibliography for more details. Let us quote some of the significant contributions to the subject. The scalar case (n = 1 or N = 1) has been intensively studied notably by Aubert and Tahraoui [4–6], Bauman and Phillips [10], Buttazzo, Ferone and Kawohl [13], Celada and Perrotta [14,15], Cellina [16,17], Cellina and Colombo [18], Cesari [20,21], Cutri [22],

62

B. Dacorogna

Dacorogna [26], Ekeland [39], Friesecke [40], Fusco, Marcellini and Ornelas [41], Giachetti and Schianchi [42], Klötzler [47], Marcellini [50–52], Mascolo [54], Mascolo and Schianchi [56,57], Monteiro Marques and Ornelas [58], Ornelas [64], Raymond [67–69], Sychev [75], Tahraoui [76,77], Treu [78] and Zagatti [80]. The vectorial case has been investigated for some special examples notably by Allaire and Francfort [3], Cellina and Zagatti [19], Dacorogna and Ribeiro [35], Dacorogna and Tanteri [37], Mascolo and Schianchi [55], Müller and Sverak [61] and Raymond [70]. A more systematic study was achieved by Dacorogna and Marcellini in [27,31,32], as well as in Dacorogna, Pisante and Ribeiro [34]. We have always considered in the present article the two important restrictions: • f does not depend on lower-order terms, i.e., f (x, u, ξ ) = f (ξ ); • the boundary datum u0 is affine, i.e., there exists ξ0 ∈ RN ×n so that Du0 = ξ0 . In the above-mentioned literature, some authors have considered either of these two more general cases. The results are then much less general and essentially apply only to the scalar case. 2. Preliminaries and notations 2.1. The different notions of convexity We start with the different definitions of convexity that we will use throughout this chapter and we refer to Dacorogna [26] for more details. D EFINITION 4. (i) A function f : RN ×n →  R = R ∪ {+∞} is said to be rank one convex if

 f λξ + (1 − λ)η  λf (ξ ) + (1 − λ)f (η)

for every λ ∈ [0, 1], ξ, η ∈ RN ×n with rank{ξ − η}  1. (ii) A Borel measurable and locally integrable function f : RN ×n → R is said to be quasiconvex if 

 1 f ξ + Dϕ(x) dx f (ξ )  meas D D

for every bounded domain D ⊂ Rn , for every ξ ∈ RN ×n and for every ϕ ∈ W01,∞ (D; RN ). (iii) A Borel measurable and locally integrable function f : RN ×n → R is said to be quasiaffine if f and −f are quasiconvex. (iv) A function f : RN ×n →  R = R ∪ {+∞} is said to be polyconvex if there exists g : Rr(n,N ) →  R convex, such that

 f (ξ ) = g T (ξ ) ,

Nonconvex problems of the calculus of variations and differential inclusions

63

where T : RN ×n → Rr(n,N ) is such that T (ξ ) = (ξ, adj2 ξ, . . . , adjn∧N ξ ). In the preceding definition, adjs ξ stands for the matrix of all s × s minors of the matrix ξ ∈ RN ×n , 2  s  n ∧ N = min{n, N }, and τ (n, N) =

n∧N 

σ (s),

s=1

where σ (s) =



N s

 N! n! n . = 2 s (s!) (N − s)!(n − s)!

R EMARK 5. (i) The concepts were introduced by Morrey [59,60], but the terminology is the one of Ball [7]; note however that Ball calls quasiaffine functions, null Lagrangians. (ii) These notions are related through the following diagram f convex ⇒

⇒

f polyconvex

f quasiconvex

⇒

f rank one convex.

In the scalar case, n = 1 or N = 1, these notions are all equivalent and reduce therefore to the usual notion of convexity. However in the vectorial case, n, N
2, these concepts are all different, meaning that there are counterexamples to all the above implications. The last counter implication being known, thanks to the celebrated example from Sverak [72], only when n
2 and N
3; the case N = 2, n
2 being still open. (iii) Note that in the case N = n = 2, the notion of polyconvexity can be read as follows:
τ (n, N) = τ (2, 2) = 5 since σ (1) = 4, σ (2) = 1, T (ξ ) = (ξ, det ξ ). (iv) Observe that, if we adopt the tensorial notation, the definition of rank one convexity can be read as follows, ϕ(t) = f (ξ + ta ⊗ b) is convex in t for every ξ ∈ RN ×n and for every a ∈ RN , b ∈ Rn , where we have denoted by 1iN

a ⊗ b = a i bα 1αn .

(v) One should also note that in the definition of quasiconvexity if the inequality holds for a given domain D ⊂ Rn , then it holds for every such domain D.

64

B. Dacorogna

(vi) If the function f : RN ×n → R, i.e., f takes only finite values, is convex or polyconvex or quasiconvex or rank one convex, then it is continuous and even locally Lipschitz. (vii) It can be shown that a quasiaffine function is necessary of the form ! " f (ξ ) = α; T (ξ ) + β for some constants α ∈ Rτ and β ∈ R and where ·; · stands for the scalar product in Rτ ; which in the case N = n = 2 reads as (α = (α11 , α12 , α21 , α22 , α5 ) ∈ R5 ) f (ξ ) =

2 

i,j =1

αij ξij + α5 det ξ + β.

(viii) An equivalent characterization of polyconvexity can be given in terms of the separation theorem (cf. Theorem 1.3 in Dacorogna [26], p. 107). A function f : RN ×n → R is polyconvex if and only if for every ξ ∈ RN ×n there exists λ = λ(ξ ) ∈ Rτ (N,n) so that ! " f (ξ + η) − f (ξ ) − λ; T (ξ + η) − T (ξ )
0 for every η ∈ RN ×n . (1)

(ix) When the function f depends on lower-order terms as in the Introduction, i.e., f : Ω × RN × RN ×n → R with f = f (x, u, ξ ), all the above notions are understood only with respect to the variable ξ, all the other variables being kept fixed. For example in the case of quasiconvex functions, one should read 

 1 f (x0 , u0 , ξ )  f x0 , u0 , ξ + Dϕ(x) dx meas D D

for every bounded domain D ⊂ Rn , for every (x0 , u0 , ξ ) ∈ Ω × RN × RN ×n and for every ϕ ∈ W01,∞ (D; RN ).

The important concept from the point of view of minimization in the calculus of variations is the notion of quasiconvexity. This condition is equivalent to the fact that the functional I, defined in the Introduction, is (sequentially) weakly lower semicontinuous in W 1,p (Ω; RN ) meaning that I (u)  lim inf I (uν ) ν→∞

for every sequence uν ⇀ u in W 1,p . Important examples of quasiconvex functions are the following. (i) The quadratic case. Let M be a symmetric matrix in R(N ×n)×(N ×n) and f (ξ ) = Mξ ; ξ , where ξ ∈ RN ×n and ·; · denotes the scalar product in RN ×n . Then f quasiconvex

⇐⇒

f rank one convex.

Nonconvex problems of the calculus of variations and differential inclusions

(ii) The Alibert–Dacorogna–Marcellini example (cf. [2]). and

65

Here we have N = n = 2

 f (ξ ) = |ξ |2 |ξ |2 − 2γ det ξ ,

where |ξ | stands for the Euclidean norm of the matrix and γ
0. Then f is convex

2√ 2, 3 γ  γp = 1,

γ  γc =

⇐⇒

⇐⇒

f is polyconvex f is quasiconvex

⇐⇒

f is rank one convex

γ  γq ,

⇐⇒

where γq > 1,

2 γ  γr = √ . 3

(iii) Let f : RN ×n → R, Φ : RN ×n → R be quasiaffine and g : R → R be such that

 f (ξ ) = g Φ(ξ )

(in particular if N = n, one can take Φ(ξ ) = det ξ ), then f polyconvex ⇐⇒

⇐⇒

f quasiconvex

f rank one convex

⇐⇒

g convex.

(iv) Let N = n + 1 and for ξ ∈ R(n+1)×n , denote

 adjn ξ = det ξˆ 1 , − det ξˆ 2 , . . . , (−1)k+1 det ξˆ k , . . . , (−1)n+2 det ξˆ n+1 ,

where ξˆ k is the n × n matrix obtained from ξ by suppressing the kth line (when ξ = Du, adjn Du represents, geometrically, the normal to the hypersurface). Let g : Rn+1 → R be such that f (ξ ) = g(adjn ξ ) then f polyconvex ⇐⇒

⇐⇒

f quasiconvex

f rank one convex

⇐⇒

g convex.

(v) Let 0  λ1 (ξ )  · · ·  λn (ξ ) denote the singular values of a matrix ξ ∈ Rn×n , which are defined as the eigenvalues of the matrix (ξ ξ ⊤ )1/2 . The functions ξ→

n  i=ν

λi (ξ )

and ξ →

n #

i=ν

λi (ξ ),

ν = 1, . . . , n,

66

B. Dacorogna

$ are respectively convex and polyconvex (note that ni=1 λi (ξ ) = | det ξ |). In particular, the function ξ → λn (ξ ) is convex and in fact is the operator norm. 2.2. Some function spaces The following notations will be used throughout. • For 1  p  ∞, we will let W 1,p (Ω; RN ) be the space of maps u : Ω ⊂ Rn → RN such that  i 1iN



 ∂u u ∈ Lp Ω, RN and Du = ∈ Lp Ω, RN ×n . ∂xj 1j n 1,p

• For 1  p < ∞, W0 (Ω; RN ) will denote the closure of C0∞ (Ω; RN ) with respect to the  · W 1,p norm. • W01,∞ (Ω; RN ) = W 1,∞ (Ω; RN ) ∩ W01,1 (Ω; RN ).  RN ) will stand for the subset of W 1,∞ (Ω; RN ) consisting of piecewise • Affpiec (Ω; affine maps. 1 (Ω;  RN ) will denote the subset of W 1,∞ (Ω; RN ) consisting of piecewise • Cpiec C 1 maps. 2.3. Statement of the problem We will be concerned with existence of minimizers for the problem




  1,∞ N , f Du(x) dx: u ∈ uξ0 + W0 Ω; R inf

(P)

Ω

where • Ω ⊂ Rn is a bounded open set with Lipschitz boundary, • u : Ω → RN and thus Du ∈ RN ×n , • f : RN ×n → R is lower semicontinuous, locally bounded and nonnegative, • ξ0 ∈ RN ×n and uξ0 is an affine map such that Duξ0 = ξ0 . The hypothesis f
0 can be replaced, with no changes, by ! " f (ξ )
α; T (ξ ) + β for every ξ ∈ RN ×n ,

for some constants α ∈ Rτ and β ∈ R and where ·; · stands for the scalar product in Rτ . This hypothesis is made to avoid to have to deal with quasiconvex envelopes Qf ≡ −∞. 3. Relaxation theorems We now present the relaxation theorem, which corresponds to the first step described in the Introduction. But before that we need to introduce the notions of envelopes correspond-

Nonconvex problems of the calculus of variations and differential inclusions

67

ing to the different concepts of convexity that we introduced in the previous section. The reference book for this part is still Dacorogna [26].

3.1. The different envelopes We now define Cf = sup{g  f : g convex}, Pf = sup{g  f : g polyconvex}, Qf = sup{g  f : g quasiconvex}, Rf = sup{g  f : g rank one convex}, they are respectively the convex, polyconvex, quasiconvex, rank one convex envelope of f. In view of the results of the previous section, we have Cf  Pf  Qf  Rf  f. As already said, we will always assume, in the sequel, that f
0. We then have the following characterizations of the different envelopes. T HEOREM 6. Let f : RN ×n →  R = R ∪ {+∞}. Part 1. Let for any integer s 

Λs = λ = (λ1 , . . . , λs ): λi
0 and

s  i=1



λi = 1 ,

then  N n+1 N n+1  ti ξi , t ∈ ΛN n+1 , ti f (ξi ): ξ = Cf (ξ ) = inf i=1

i=1

 τ +1 τ +1   ti T (ξi ), t ∈ Λτ +1 . ti f (ξi ): T (ξ ) = Pf (ξ ) = inf i=1

i=1

Part 2. Let R0 f = f and define inductively for i an integer,  Ri+1 f (ξ ) = inf tRi f (ξ1 ) + (1 − t)Ri f (ξ2 ): t ∈ [0, 1], ξ = tξ1 + (1 − t)ξ2 , rank{ξ1 − ξ2 } = 1 ,

68

B. Dacorogna

then Rf (ξ ) = inf Ri f (ξ ). i∈N

T HEOREM 7 (Dacorogna formula). If f : RN ×n → R is locally bounded and Borel measurable then






 1 Qf (ξ ) = inf f ξ + Dϕ(x) dx: ϕ ∈ W01,∞ Ω; RN , meas Ω Ω where Ω ⊂ Rn is a bounded domain. In particular, the infimum is independent of the choice of the domain. R EMARK 8. (i) The representation formula for Cf is standard and follows from Carathéodory theorem. The inductive way of representing Rf was found by Kohn and Strang [48]. The formulas for Pf and Qf (and a similar to that of Kohn and Strang for Rf ) were established by Dacorogna (cf. [26]). (ii) Using the separation theorems one can establish other formulas for Cf and Pf, cf. [26].

3.2. Some examples We now discuss some examples that will be used in Section 7. We start with the following theorem established by Dacorogna [26] (cf. also [23] and [24]). T HEOREM 9. Part 1. Let f : RN ×n → R, Φ : RN ×n → R be quasiaffine and g : R → R be such that

 f (ξ ) = g Φ(ξ )

(in particular if N = n, one can take Φ(ξ ) = det ξ ), then

 Pf (ξ ) = Qf (ξ ) = Rf (ξ ) = Cg Φ(ξ ) .

Part 2. Let N = n + 1, f : RN ×n → R and g : Rn+1 → R be such that f (ξ ) = g(adjn ξ ), then Pf (ξ ) = Qf (ξ ) = Rf (ξ ) = Cg(adjn ξ ).

Nonconvex problems of the calculus of variations and differential inclusions

69

The next result, established by Dacorogna, Pisante and Ribeiro [34], concerns functions depending on singular values. We let N = n and we denote by λ1 (ξ ), . . . , λn (ξ ) the singular values of ξ ∈ Rn×n with 0  λ1 (ξ )  · · ·  λn (ξ ) (which are the eigenvalues of the matrix (ξ ξ ⊤ )1/2 ) and by Q the set  Q = x = (x2 , . . . , xn−1 ) ∈ Rn−2 : 0  x2  · · ·  xn−1

which is the natural set, where to consider (λ2 (ξ ), . . . , λn−1 (ξ )) for ξ ∈ Rn×n . T HEOREM 10. Let g : Q × R → R, g = g(x, s), be a function such that x → g(x, s) is continuous and bounded from below for all s ∈ R. Let f : Rn×n → R be defined by

then

 f (ξ ) = g λ2 (ξ ), . . . , λn−1 (ξ ), det ξ , Pf (ξ ) = Qf (ξ ) = Rf (ξ ) = Ch(det ξ ),

where h : R → R is given by h(s) = infx∈Q g(x, s). R EMARK 11. We remark that if some dependence on λ1 or λn is allowed, then no simple and general expression for the envelopes is known; see [34], when there is dependence on λ1 , and Theorem 3.5 by Buttazzo, Dacorogna and Gangbo [12], when there is dependence on λn . The next result concerns the Saint Venant–Kirchhoff energy function, which is particularly important in nonlinear elasticity. The function, up to rescaling, is given by, ν ∈ (0, 1/2) being a parameter,  2 f (ξ ) = ξ ξ ⊤ − I  +

2 ν 2 |ξ | − n 1 − 2ν

or in terms of the singular values, 0  λ1 (ξ )  · · ·  λn (ξ ), of ξ ∈ Rn×n , n 

2 2 f (ξ ) = λi − 1 + i=1

2  n  ν 2 λi − n . 1 − 2ν i=1

Le Dret and Raoult [49] have computed the quasiconvex envelope when n = 2 or n = 3 and they have shown the following. T HEOREM 12. If n = 2 or n = 3, then Qf (ξ ) = Cf (ξ ).

70

B. Dacorogna

When n = 2 it is given by

where

   f (ξ ) 2 1 Cf (ξ ) = Pf (ξ ) = Qf (ξ ) = Rf (ξ ) = 1−ν λ22 − 1   0

if ξ ∈ / D 1 ∪ D2 , if ξ ∈ D2 , if ξ ∈ D1 ,

 % &2 % &2 D1 = ξ ∈ R2×2 : (1 − ν) λ1 (ξ ) + ν λ2 (ξ ) < 1 and λ2 (ξ ) < 1  = ξ ∈ R2×2 : λ1 (ξ )  λ2 (ξ ) < 1 ,  % &2 % &2 D2 = ξ ∈ R2×2 : (1 − ν) λ1 (ξ ) + ν λ2 (ξ ) < 1 and λ2 (ξ )
1 .

The last example is related to a problem of optimal design and has been studied by Kohn and Strang [48]. T HEOREM 13. Let n = N = 2 and f (ξ ) =




1 + |ξ |2 0

if ξ = 0,

if ξ = 0.

Then Pf = Qf = Rf and Qf (ξ ) =



1 + |ξ |2 if |ξ |2 + 2| det ξ |
1, 1/2

2 2 |ξ | + 2| det ξ | − 2| det ξ | if |ξ |2 + 2| det ξ | < 1.

R EMARK 14. The above result is still valid when N
3, it suffices to replace det ξ by N adj2 ξ ∈ R( 2 ) . 3.3. The main theorem We now turn our attention to the relaxation theorem. We recall our minimization problem




  1,p inf I (u) = f Du(x) dx: u ∈ u0 + W0 Ω; RN ,

(P)

Ω

where 1  p  ∞. We define the relaxed problem associated to (P) to be


 

 1,p Qf Du(x) dx: u ∈ u0 + W0 Ω; RN . inf I(u) = Ω

(QP)

Nonconvex problems of the calculus of variations and differential inclusions

71

T HEOREM 15 (Relaxation theorem). Let Ω ⊂ Rn be a bounded open set. Let f : RN ×n → R be Borel measurable and nonnegative satisfying, for 1  p < ∞,

 0  f (ξ )  α1 1 + |ξ |p for every ξ ∈ RN ×n ,

(2)

where α1 > 0 is a constant and for p = ∞ it is assumed that f is locally bounded. Let Qf = sup{g  f : g quasiconvex} be the quasiconvex envelope of f. Then inf(P) = inf(QP). More precisely, for every u ∈ W 1,p (Ω; RN ), there exists a sequence {uν }∞ ν=1 ⊂ u0 + 1,p W0 (Ω; RN ) such that 

Ω

 f Duν (x) dx →



Ω

 Qf Du(x) dx

as ν → ∞.

R EMARK 16. (i) If we add in the theorem a coercivity condition



 α2 −1 + |ξ |p  f (ξ )  α1 1 + |ξ |p ,

where α2 > 0 and p > 1, we can infer that (QP) has a minimizer and that the sequence {uν }∞ ν=1 further satisfies uν ⇀ u

 in W 1,p Ω; RN as ν → ∞.

(ii) The theorem remains also valid if the function f depends on lower-order terms, i.e., f = f (x, u, ξ ). The quasiconvex envelope is then to be understood as the quasiconvex envelope only with respect to the variable ξ, the other variables (x, u) being kept fixed. P ROOF OF T HEOREM 15. We divide the proof into two steps. Step 1. We start with an approximation of the given function u. Let ε > 0 be arbitrary, we can then find disjoint open sets Ω1 , . . . , Ωk ⊂ Ω, ξ1 , . . . , ξk ∈ RN ×n , γ independent of ε 1,p and v ∈ u + W0 (Ω; RN ) such that  % & k   meas Ω − i=1 Ωi  ε, uW 1,p , vW 1,p  γ , u − vW 1,1  ε,   Dv(x) = ξi if x ∈ Ωi .

(3)

72

B. Dacorogna

By taking ε smaller if necessary we can also assume, using the continuity of Qf and the growth condition on f, that   

 Qf Du(x) − Qf Dv(x)  dx  ε, (4) Ω  % 

& 0 f Dv(x) − Qf Dv(x) dx  ε. (5)  k i=1 Ωi

Ω−

Indeed let us discuss the case 1  p < ∞, the case p = ∞ being easy. As is well known (cf. Lemma 2.2 in [26], p. 156) any quasiconvex function is locally Lipschitz continuous and if it satisfies (2), then there exists β > 0 such that  

 Qf (Du) − Qf (Dv)  β 1 + |Du|p−1 + |Dv|p−1 |Du − Dv|. Using the Hölder inequality we obtain    Qf (Du) − Qf (Dv) dx Ω ((p−1)/p ' (1/p ' % & p−1 p−1 p/(p−1) p 1 + |Du| + |Dv| |Du − Dv| β Ω

Ω

and (4) follows therefore from (3). The inequality (5) follows from (3) and a classical property of the integrals (cf. Lemma 1.4 in [26], p. 19). Step 2. Now use Theorem 7 on every Ωi to find ϕi ∈ W01,∞ (Ωi ; RN ) 

 f ξi + Dϕi (x) dx Ωi 

 1
Qf (ξi )
−ε + f ξi + Dϕi (x) dx. meas Ωi Ωi

1 meas Ωi

Setting w(x) =




v(x) + ϕi (x) if x ∈ Ωi , i = 1, . . . , k,  v(x) if x ∈ Ω − ki=1 Ωi , 1,p

we get that w ∈ u + W0 (Ω; RN ) and (using (5)) 0 0 =



k





i=1 Ωi

) k + * % 

& f Dw(x) − Qf Dv(x) dx  ε meas Ωi ,

 Ω− ki=1 Ωi

 Ω− ki=1 Ωi

i=1

% 

& f Dw(x) − Qf Dv(x) dx

% 

& f Dv(x) − Qf Dv(x) dx  ε.

Nonconvex problems of the calculus of variations and differential inclusions

73

In other words, combining these inequalities, we have proved that  % 

& 0 f Dw(x) − Qf Dv(x) dx  ε(1 + meas Ω). Ω

Invoking (4), we find    % 

&   f Dw(x) − Qf Du(x) dx   ε(2 + meas Ω).  Ω

Setting ε = 1/ν with ν ∈ N and uν = w, we have indeed obtained the theorem.



We now discuss the history of this theorem (for precise references see [26]). In the case N = n = 1, this result has been proved by L.C. Young and then generalized by others to the scalar case, N = 1 or n = 1, notably by Berliochi and Lasry, Ekeland, Ioffe and Tihomirov and Marcellini and Sbordone. Note that in this context Qf = Cf = f ∗∗ , where Cf is the usual convex envelope of f. The problem (QP) can then be rewritten as


 

 1,p ∗∗ N ∗∗ . (P∗∗ ) inf I (u) = f Du(x) dx: u ∈ u0 + W0 Ω; R Ω

The result for the vectorial case (i.e., N, n > 1, recall also that, in general, we now have Qf > Cf ) was established by Dacorogna in [25]. Following a different approach it was later also proved by Acerbi and Fusco [1]. In the present context the equivalence between (QP) and (P∗∗ ) is not any more valid, one has in general

 inf(P) = inf(QP) > inf P∗∗ .

The inequality is, in general, strict as in the simple example where N = n
2 and f (ξ ) = (det ξ )2 . We indeed have f (ξ ) = Qf (ξ ) = (det ξ )2

and f ∗∗ (ξ ) ≡ 0.

Therefore if det Du0 > 0, then, using the Jensen inequality, we have inf(P) = inf(QP) 
meas Ω

1 meas Ω

∗∗  > 0 = inf P .



det Du0 (x) dx Ω

2

Closely related to this approach is the notion of parametrized or Young measure, that we do not discuss here.

74

B. Dacorogna

4. Implicit partial differential equations 4.1. Introduction We now discuss the existence of solutions, u ∈ W 1,∞ (Ω; RN ), for the Dirichlet problem involving differential inclusions of the form


Du(x) ∈ E u(x) = ϕ(x),

a.e. in Ω, x ∈ ∂Ω,

where ϕ is a given function and E ⊂ RN ×n is a given compact set. To relate this study with what we said in Section 1, one should imagine that  E = ξ ∈ RN ×n : Qf (ξ ) = f (ξ )

and therefore the differential inclusion is equivalent to the implicit partial differential equation



 Qf Du(x) = f Du(x)

a.e. x ∈ Ω.

In the scalar case (n = 1 or N = 1) a sufficient condition for solving the problem is Dϕ(x) ∈ E ∪ int co E

a.e. in Ω,

where int co E stands for the interior of the convex hull of E. This fact was observed by several authors, with different proofs and different levels of generality; notably in [11,17, 28,29,31,38,40]. It should be noted that this sufficient condition is also necessary, when properly reformulated. When turning to the vectorial case (n, N
2) the problem becomes considerably harder and no result with such a degree of elegancy and generality is available. The first general results were obtained by Dacorogna and Marcellini (see the References, in particular, [31]). At the same time Müller and Sverak [61] introduced the method of convex integration of Gromov in this framework, obtaining also similar existence results.

4.2. The different convex hulls We recall the main notations that we will use throughout the present section and we refer, if necessary, for more details to Dacorogna and Marcellini [31]. Classically the convex hull of a given set E is the smallest convex set that contains E and it is denoted by co E. We will now do the same with the other notions of convexity that we have seen earlier. This is not as straightforward as it may seem and there is not a general agreement on the exact definitions. We will not enter in abstract considerations and we will use as definition of the different hulls a consequence of these abstract definitions.

Nonconvex problems of the calculus of variations and differential inclusions

75

N OTATION 17. We let, for E ⊂ RN ×n ,  E = f : RN ×n →  F R = R ∪ {+∞}: f |E  0 ,  FE = f : RN ×n → R: f |E  0 .

We then have respectively the convex, polyconvex, rank one convex and (closure of the) quasiconvex hull defined by  E , co E = ξ ∈ RN ×n : f (ξ )  0 for every convex f ∈ F  E , Pco E = ξ ∈ RN ×n : f (ξ )  0 for every polyconvex f ∈ F  Rco E = ξ ∈ RN ×n : f (ξ )  0 for every rank one convex f ∈ F E ,  QcoE = ξ ∈ RN ×n : f (ξ )  0 for every quasiconvex f ∈ FE .

E by FE in the definitions of co E and Pco E We should point out that by replacing F we get their closures denoted by co E and Pco E. However if we do so in the definition of Rco E we get a larger set than the closure of Rco E. We should also draw the attention that some authors call the set  E ξ ∈ RN ×n : f (ξ )  0 for every rank one convex f ∈ F the lamination convex hull, while they reserve the name of rank one convex hull to the set  ξ ∈ RN ×n : f (ξ )  0 for every rank one convex f ∈ FE .

We think however that our terminology is more consistent with the classical definition of convex hull. In general we have, for any set E ⊂ RN ×n , E ⊂ Rco E ⊂ Pco E ⊂ co E, E ⊂ RcoE ⊂ QcoE ⊂ PcoE ⊂ coE. 4.3. Some examples of convex hulls We now give several examples that will be used in the applications of Sections 6 and 7. Let us start with the scalar case. E XAMPLE 18 (Convex Hamiltonian). Let F : Rn → R be convex and let  E = ξ ∈ Rn : F (ξ ) = 0 ,

then

 co E = ξ ∈ Rn : F (ξ )  0 .

76

B. Dacorogna

E XAMPLE 19 (Nonconvex Hamiltonian). Consider, for ξ ∈ Rn , the nonconvex Hamiltonian F (ξ ) =

n  % 2 &2 (ξi ) − 1 i=1

and

then

 E = ξ ∈ Rn : F (ξ ) = 0 , co E = [−1, 1]n .

We now turn to some examples in the vectorial case. The following result is due to Dacorogna and Tanteri (cf. [36] and also [31]), it concerns singular values. We recall that we denote by 0  λ1 (ξ )  · · ·  λn (ξ ) the singular values of a matrix ξ ∈ Rn×n , which are defined as the eigenvalues of the matrix (ξ ξ ⊤ )1/2 . T HEOREM 20. Let  E = ξ ∈ Rn×n : λi (ξ ) = ai , i = 1, . . . , n ,

where 0 < a1  · · ·  an . The following equalities then hold 

co E = ξ ∈ R

n×n

:

n 

λi (ξ ) 

i=ν



n  i=ν

Pco E = QcoE = Rco E = ξ ∈ R 

int Rco E = ξ ∈ R

n×n

:

n #

i=ν

ai , ν = 1, . . . , n ,

n×n

λi (ξ )
0 be arbitrary. We wish to We now show that V is dense in V k find v ∈ V so that u − vL∞  ε. We recall (cf. Section 4.6) that ωD (α) = lim

sup

δ→0 v,w∈B∞ (α,δ)

Dv − DwL1 (Ω) ,

where  : u − αL∞ < δ . B∞ (α, δ) = u ∈ V

80

B. Dacorogna

, a point of continuity of the operator D, so that • We start by finding α ∈ V ε u − αL∞  . 3 This is always possible by virtue of Corollary 40. In particular, we have that the oscillation ωD (α) of the gradient operator at α is zero.  by β ∈ V so that • We next approximate α ∈ V β − αL∞ 

ε 3

and ωD (β)
0 the set  ε : ωD (u) < ε ΩD := u ∈ V

. is open in V • Finally we use the relaxation property on every piece where Dβ is constant and we then construct v ∈ V , by patching all the pieces together, such that ε β − vL∞  , 3

ωD (v)
0 so that Dv − DψL1 

1 2k

and hence,  



 1 dist Dψ(x); E dx  dist Dv(x); E dx + Dv − DψL1 < k Ω Ω for every ψ ∈ B∞ (v, δ); which implies that v ∈ V k . Combining these three facts we have indeed obtained the desired density result.



To conclude this subsection we give a sufficient condition that ensures the relaxation property. In concrete examples this condition is usually much easier to check than the relaxation property. We start with a definition. D EFINITION 27 (Approximation property). Let E ⊂ K(E) ⊂ RN ×n . The sets E and K(E) are said to have the approximation property if there exists a family of closed sets Eδ and K(Eδ ), δ > 0, such that (1) Eδ ⊂ K(Eδ ) ⊂ int K(E) for every δ > 0; (2) for every ε > 0 there exists δ0 = δ0 (ε) > 0 such that dist(η; E)  ε for every η ∈ Eδ and δ ∈ [0, δ0 ];

Nonconvex problems of the calculus of variations and differential inclusions

81

(3) if η ∈ int K(E), then η ∈ K(Eδ ) for every δ > 0 sufficiently small. We therefore have the following theorem (cf. Theorem 6.14 in [31], and for a slightly more flexible one see Theorem 6.15). T HEOREM 28. Let E ⊂ RN ×n be compact and Rco E has the approximation property with K(Eδ ) = Rco Eδ , then int Rco E has the relaxation property with respect to E. 4.5. Some examples of existence of solutions We now give several examples of existence theorems that follow from the abstract ones. The first one concerns the scalar case, where we can even get sharper results (cf. [11,17, 28,29,31,38,40]). T HEOREM 29. Let Ω ⊂ Rn be a bounded open set and E ⊂ Rn . Let ϕ ∈ W 1,∞ (Ω) satisfy Dϕ(x) ∈ E ∪ int co E

a.e. x ∈ Ω

(6)

(where int co E stands for the interior of the convex hull of E), then there exists u ∈ ϕ + W01,∞ (Ω) such that Du(x) ∈ E

a.e. x ∈ Ω.

(7)

R EMARK 30. The theorem is in fact much less restrictive than the abstract one, here we do not need, for example, E to be compact. For a proof we refer to Dacorogna and Marcellini [31]. We now show that (6) is in fact also a necessary condition, at least when ϕ is affine, for the general case see Section 2.4 in [31]. For the affine case the result is implicit in the above-mentioned articles, but we follow here Bandyopadhyay, Barroso, Dacorogna and Matias [9]. T HEOREM 31. Let Ω ⊂ Rn be a bounded open set, E ⊂ Rn , ξ0 ∈ Rn and u ∈ uξ0 + W01,∞ (Ω) (uξ0 being such that Duξ0 = ξ0 ) so that Du(x) ∈ E

a.e. x ∈ Ω,

then ξ0 ∈ E ∪ int co E. P ROOF. Assume that ξ0 ∈ / E, otherwise nothing is to be proved. It is easy to see that, by the Jensen inequality and since Du(x) ∈ E,  1 ξ0 = Du(x) dx ∈ coE. meas Ω Ω

82

B. Dacorogna

Let us show that we cannot have ξ0 ∈ ∂(coE). If we can prove this, we will deduce that ξ0 ∈ int coE. Since int coE = int co E (cf. Theorem 6.3 in Rockafellar [71]), we will have the result. If ξ0 ∈ ∂(coE), we find from the separation theorem that there exists α ∈ Rn , α = 0, such that α; z − ξ0 
0 ∀z ∈ coE. We therefore have that ! " α; Du(x) − ξ0
0 a.e. x ∈ Ω.

Recalling that u ∈ uξ0 + W01,∞ (Ω), we find that 

Ω

! " α; Du(x) − ξ0 dx = 0

which coupled with the above inequality leads to ! " α; Du(x) − ξ0 = 0 a.e. x ∈ Ω.

Applying Lemma 57, we get that u ≡ uξ0 and hence ξ0 ∈ E, a contradiction with the hypothesis made at the beginning of the proof. Therefore ξ0 ∈ / ∂(coE) as claimed and hence the theorem is proved.  Theorem 29 applies to the following case. C OROLLARY 32 (Convex Hamiltonian). Let Ω ⊂ Rn be a bounded open set and F : Rn → R be convex and such that lim|ξ |→∞ F (ξ ) = +∞. Let ϕ ∈ W 1,∞ (Ω) be such that

 F Dϕ(x)  0 a.e. x ∈ Ω.

Then there exists u ∈ ϕ + W01,∞ (Ω) such that

 F Du(x) = 0 a.e. x ∈ Ω.

The next one deals with the singular values case that we have encountered in Section 4.3. The next theorem is due to Dacorogna and Ribeiro [35]. T HEOREM 33 (Singular values). Let Ω ⊂ Rn be a bounded open set, α < β and 0 < a2  · · ·  an be such that n #  max |α|, |β|  a2 ai . i=2

Nonconvex problems of the calculus of variations and differential inclusions

83

1 (Ω;  Rn ) be such that, for almost every x ∈ Ω, Let ϕ ∈ Cpiec n #

α < det Dϕ(x) < β,

i=ν

n

 # ai , λi Dϕ(x) < i=ν

ν = 2, . . . , n.

Then there exists u ∈ ϕ + W01,∞ (Ω; Rn ) so that, for almost every x ∈ Ω,

 det Du(x) ∈ {α, β}, λν Du(x) = aν , ν = 2, . . . , n. R EMARK 34. (i) If α = −β < 0 and if we set a1 = β

)

n #

ai

i=2

+−1

,

we recover the result of Dacorogna and Marcellini [31], namely that if n #

i=ν

n

 # λi Dϕ(x) < ai ,

ν = 1, . . . , n,

i=ν

then there exists u ∈ ϕ + W01,∞ (Ω; Rn ) so that λν (Du) = aν ,

ν = 1, . . . , n, a.e. in Ω.

(ii) If α = β = 0 we can also prove, as in Dacorogna and Tanteri [37], that if det Dϕ(x) = α,

n #

i=ν

n

 # ai , λi Dϕ(x) < i=ν

ν = 2, . . . , n,

then there exists u ∈ ϕ + W01,∞ (Ω; Rn ) so that λν (Du) = aν ,

ν = 2, . . . , n,

and

det Du = α

a.e. in Ω.

4.6. Appendix In this appendix we recall some well-known facts about the so-called functions of first class in the sense of Baire, with particular interest in their application to the gradient operator. We start recalling some definitions. D EFINITION 35. Let X, Y be metric spaces and f : X → Y . We define the oscillation of f at x0 ∈ X as

 ωf (x0 ) = lim sup dY f (y), f (x) , δ→0 x,y∈BX (x0 ,δ)

84

B. Dacorogna

where BX (x0 , δ) := {x ∈ X: dX (x, x0 ) < δ} is the open ball centered at x0 and dX , dY are the metric on the spaces X and Y , respectively. D EFINITION 36. A function f is said to be of first class (in the sense of Baire) if it can be represented as the pointwise limit of an everywhere convergent sequence of continuous functions. In the next proposition we recall some elementary properties of the oscillation function ωf . P ROPOSITION 37. Let X, Y be metric spaces, and f : X → Y. (i) f is continuous at x0 ∈ X if and only if ωf (x0 ) = 0. (ii) The set Ωfε := {x ∈ X: ωf (x) < ε} is an open set in X. Using the notion of oscillation and Proposition 37 we can write the set Df of all points at which a given function f is discontinuous as an Fσ set as follows

∞
* 1 Df = x ∈ X: ωf (x)
. n

(8)

n=1

We therefore have the following Baire theorem for functions of first class (for a proof see Theorem 7.3 in Oxtoby [65], Yosida [79], p. 12, or Dacorogna and Pisante [33]). T HEOREM 38. Let X, Y be metric spaces let X be complete and f : X → Y . If f is a function of first class, then Df is a set of first category. R EMARK 39. From Theorem 38 and the Baire category theorem follows in particular that the set of points of continuity of a function of first class from a complete metric space X to any metric space Y , i.e., the set Dfc complement of Df , is a dense Gδ set. Indeed for any ε > 0, the set  Ωfε := x ∈ X: ωf (x) < ε

is open and dense in X.

In the proof of our main theorem we have used Theorem 38 applied to the following, quite surprising, special case of function of first class. This result was observed by Kirchheim [46] for complete sets of Lipschitz functions and the same argument gives in fact the result for general complete subsets W 1,∞ (Ω) functions. C OROLLARY 40. Let Ω ⊂ Rn be a bounded open set and let V ⊂ W 1,∞ (Ω) be a nonempty complete space with respect to the L∞ metric. Then the gradient operator D : V → Lp (Ω; Rn ) is a function of first class for any 1  p < ∞.

Nonconvex problems of the calculus of variations and differential inclusions

85

P ROOF. For h = 0, we let



 D h = D1h , . . . , Dnh : V → Lp Ω; Rn

be defined, for every u ∈ V and x ∈ Ω, by Dih u(x) =



u(x+hei )−u(x) h

0



if dist x, Ω c > |h|, elsewhere,

for i = 1, . . . , n, where e1 , . . . , en stand for the vectors from the Euclidean basis. The claim will follow once we will have proved that for any fixed h the operator D h is continuous and that, for any sequence h → 0,   lim Dih u − Di uLp (Ω) = 0

h→0

for any i = 1, . . . , n, u ∈ V . The continuity of D h follows easily by observing that for every i = 1, . . . , n, ε > 0 and u, v ∈ V we have that   h D u − D h v  p i i L (Ω)  1/p   1 u(x) − v(x) + u(x + hei ) − v(x + hei )p dx  |h| Ωh 

2(meas Ω)1/p u − vL∞ (Ω) , |h|

where Ωh = {x ∈ Ω: dist(x, Ω c ) > |h|}. For the second claim we start observing that, for any x ∈ Ωh and for any u ∈ V , we have   u(x + hei ) − u(x)  Di uL∞ (Ω) |h|.

This implies that

 h  D u ∞  Di uL∞ (Ω) < +∞. i L (Ω)

Moreover, by Rademacher theorem, for any sequence h → 0, lim Dih u(x) = Di u(x) a.e. x ∈ Ω.

h→0

The result follows by Lebesgue dominated convergence theorem.



86

B. Dacorogna

5. Existence of minimizers 5.1. Introduction We now discuss the existence of minimizers for the problem






 inf f Du(x) dx: u ∈ uξ0 + W01,∞ Ω; RN ,

(P)

Ω

where Ω ⊂ Rn is a bounded open set with Lipschitz boundary, u : Ω → RN , f : RN ×n → R is lower semicontinuous, locally bounded and nonnegative and uξ0 is a given affine map (i.e., Duξ0 = ξ0 , where ξ0 ∈ RN ×n is a fixed matrix). If the function f is quasiconvex, i.e., 

 f ξ + Dϕ(x) dx
f (ξ ) meas(U ) U

for every bounded domain U ⊂ Rn , ξ ∈ RN ×n and ϕ ∈ W01,∞ (U ; RN ), then the problem (P) trivially has uξ0 as a minimizer. We also recall that in the scalar case (n = 1 or N = 1), quasiconvexity and ordinary convexity are equivalent. We now study the case where f fails to be quasiconvex. The first step in dealing with such problems is the relaxation theorem (cf. Theorem 15). It has as a direct consequence (cf. Theorem 41) that (P) has a solution u¯ ∈ uξ0 + W01,∞ (Ω; RN ) if and only if



 f D u(x) ¯ = Qf D u(x) ¯ a.e. x ∈ Ω, 

 Qf D u(x) ¯ dx = Qf (ξ0 ) meas Ω, Ω

where Qf is the quasiconvex envelope of f , namely Qf = sup{g  f : g quasiconvex}. The problem is then to discuss the existence or nonexistence of a u¯ satisfying the two equations. The two equations are not really of the same nature. The first one is what we called in Section 4 an implicit partial differential equation. The second one is more geometric in nature and has to do with some “quasiaffinity” of the quasiconvex envelope Qf. In the present section we will discuss some abstract necessary and sufficient conditions for the existence of minimizers for (P) and in Sections 6 and 7 we will see several examples. We will follow in the present section the approach of Dacorogna, Pisante and Ribeiro [34].

5.2. Sufficient conditions With the help of the relaxation theorem and of Theorem 25, we are now in a position to discuss some existence results for the problem (P). The following theorem (cf. [27]) is

Nonconvex problems of the calculus of variations and differential inclusions

87

elementary and gives a necessary and sufficient condition for existence of minima. It will be crucial in several of our arguments. T HEOREM 41. Let Ω, f and uξ0 be as above, in particular, Duξ0 = ξ0 . The problem (P) has a solution if and only if there exists u¯ ∈ uξ0 + W01,∞ (Ω; RN ) such that



 f D u(x) ¯ = Qf D u(x) ¯ a.e. x ∈ Ω, 

 Qf D u(x) ¯ dx = Qf (ξ0 ) meas Ω.

(9) (10)

Ω

P ROOF. By the relaxation theorem and since uξ0 is affine, we have inf(P) = inf(QP) = Qf (ξ0 ) meas Ω. Moreover, since we always have f
Qf and we have a solution of (9) satisfying (10), we get that u¯ is a solution of (P). The fact that (9) and (10) are necessary for the existence of a minimum for (P) follows in the same way.  The previous theorem explains why the set  K = ξ ∈ RN ×n : Qf (ξ ) < f (ξ )

plays a central role in the existence theorems that follow. In order to ensure (9) we will have to consider differential inclusions of the form studied in the previous section, namely: find u¯ ∈ uξ0 + W01,∞ (Ω; RN ) such that D u(x) ¯ ∈ ∂K

a.e. x ∈ Ω.

In order to deal with the second condition (10) we will have to impose some hypotheses of the type “Qf is quasiaffine on K”. The main abstract theorem is the following one. T HEOREM 42. Let Ω ⊂ Rn be a bounded open set, ξ0 ∈ RN ×n , f : RN ×n → R, a lower semicontinuous, locally bounded and nonnegative function and let  K = η ∈ RN ×n : Qf (η) < f (η) .

Assume that there exists K0 ⊂ K such that • ξ0 ∈ K0 , 0 ∩ ∂K, • K0 is bounded and has the relaxation property with respect to K 0 . • Qf is quasiaffine on K Let uξ0 (x) = ξ0 x. Then the problem




  1,∞ N inf I (u) = f Du(x) dx: u ∈ uξ0 + W0 Ω; R Ω

(P)

88

B. Dacorogna

has a solution u¯ ∈ uξ0 + W01,∞ (Ω; RN ). R EMARK 43. (i) Although this theorem applies only to functions f that takes only finite R = R ∪ {+∞}. values, it can sometimes be extended to functions f : RN ×n →  (ii) The last hypothesis in the theorem means that 

 Qf ξ + Dϕ(x) dx = Qf (ξ ) meas Ω Ω

0 , every ϕ ∈ W 1,∞ (Ω; RN ) with for every ξ ∈ K 0 0 ξ + Dϕ(x) ∈ K

a.e. in Ω.

P ROOF OF T HEOREM 41. Since ξ0 ∈ K0 and K0 is bounded and has the relaxation 0 ∩ ∂K, we can find, appealing to Theorem 25, a map u¯ ∈ property with respect to K 1,∞ N uξ0 + W0 (Ω; R ) satisfying 0 ∩ ∂K D u¯ ∈ K

a.e. in Ω,

0 , which means that (9) of Theorem 41 is satisfied. Moreover, since Qf is quasiaffine on K we have that (10) of Theorem 41 holds and thus the claim.  The second hypothesis in the theorem is clearly the most difficult to verify, nevertheless there are some cases when it is automatically satisfied. For example, if K is bounded, we can take K0 = K. We will see that, in many applications, the set K turns out to be unbounded and in order to apply Theorem 42 we need to find some weaker conditions on K that guarantees the existence of a subset K0 of K satisfying the requested properties. With this aim in mind we give the following notation and definition. N OTATION 44. Let K ⊂ RN ×n be open and λ ∈ RN ×n . (i) For ξ ∈ K, we denote by LK (ξ, λ) the largest segment of the form [ξ + tλ, ξ + sλ], t < 0 < s, so that (ξ + tλ, ξ + sλ) ⊂ K. (ii) If LK (ξ, λ) is bounded, we denote by t− (ξ ) < 0 < t+ (ξ ) the elements so that LK (ξ, λ) = [ξ + t− λ, ξ + t+ λ]. They therefore satisfy ξ + t± λ ∈ ∂K

and ξ + tλ ∈ K

∀t ∈ (t− , t+ ).

(iii) If H ⊂ K, we let LK (H, λ) =

*

LK (ξ, λ).

ξ ∈H

D EFINITION 45. Let K ⊂ RN ×n be open, ξ0 ∈ K and λ ∈ RN ×n . (i) We say that K is bounded at ξ0 in the direction λ if LK (ξ0 , λ) is bounded.

Nonconvex problems of the calculus of variations and differential inclusions

89

(ii) We say that K is stably bounded at ξ0 in the rank-one direction λ = α ⊗ β (with α ∈ RN and β ∈ Rn ) if there exists ε > 0 so that LK (ξ0 + α ⊗ Bε , λ) is bounded, where we have denoted by  ξ0 + α ⊗ Bε = ξ ∈ RN ×n : ξ = ξ0 + α ⊗ b with |b| < ε .

Clearly a bounded open set K is bounded at every point ξ ∈ K and in any direction λ and consequently it is also stably bounded. We now give an example of a globally unbounded set which is bounded in certain directions. E XAMPLE 46. Let N = n = 2 and

 K = ξ ∈ R2×2 : α < det ξ < β .

The set K is clearly unbounded. (i) If ξ0 = I then K is bounded, and even stably bounded, at ξ0 , in a direction of rank one, for example, with λ=



1 0 0 0



or λ =



0 0 . 0 1

(ii) However if ξ0 = 0, then K is unbounded in any rank one direction, but is bounded in any rank two direction. In the following result we deal with sets K that are bounded in a rank-one direction only. This corollary says, roughly speaking, that if K is bounded at ξ0 in a rank-one direction λ and this boundedness (in the same direction) is preserved under small perturbations of ξ0 along rank one λ-compatible directions, then we can ensure the relaxation property required in the main existence theorem. C OROLLARY 47. Let Ω ⊂ Rn be a bounded open set, f : RN ×n → R a lower semicontinuous function, locally bounded and nonnegative and let ξ0 ∈ K, where  K = ξ ∈ RN ×n : Qf (ξ ) < f (ξ ) .

If there exist a rank-one direction λ ∈ RN ×n such that (i) K is stably bounded at ξ0 in the direction λ = α ⊗ β, ε , λ), then the problem (ii) Qf is quasiaffine on the set (cf. Definition 45) LK (ξ0 + α ⊗ B


 

 1,∞ N Ω; R inf I (u) = f Du(x) dx: u ∈ uξ0 + W0 Ω

has a solution u¯ ∈ uξ0 + W01,∞ (Ω; RN ).

(P)

90

B. Dacorogna

P ROOF. We divide the proof into two steps. Step 1. Assume that |β| = 1, otherwise replace it by β/|β|, and let βk ∈ Rn , k
n, with |βk | = 1, be such that  0 ∈ H := int co{β, −β, β3 , . . . , βk } ⊂ B1 (0) = x ∈ Rn : |x| < 1 . Let then, for ε > 0 as in the hypothesis, %

& , λ . K0 := (ξ0 + α ⊗ εH ) ∪ ∂K ∩ LK ξ0 + α ⊗ ε H

We therefore have that ξ0 ∈ K0 and, by hypothesis, that K0 is bounded, since

 0 ⊂ LK ξ0 + α ⊗ B ε , λ . K0 ⊂ K

Furthermore we have

 0 ∩ ∂K = ∂K ∩ LK ξ0 + α ⊗ ε H , λ . K

In order to deduce the corollary from Theorem 42, we only need to show that K0 has the 0 ∩ ∂K. This will be achieved in the next step. relaxation property with respect to K 0 ∩ ∂K. Let Step 2. We now prove that K0 has the relaxation property with respect to K  RN ) so that ξ ∈ K0 and let us find a sequence uν ∈ Affpiec (Ω;



 0 ∩ ∂K ∪ K0 a.e. in Ω, Duν (x) ∈ K uν ∈ uξ + W01,∞ Ω; RN ,  (11)

 ∗ 0 ∩ ∂K dx → 0 as ν → ∞. uν ⇀ uξ in W 1,∞ , dist Duν (x); K Ω

, λ), nothing is to be proved; so we assume that ξ ∈ ξ0 + If ξ ∈ ∂K ∩ LK (ξ0 + α ⊗ ε H α ⊗ εH. By hypothesis (i), we can find t− (ξ ) < 0 < t+ (ξ ) so that ξ± := ξ + t± λ ∈ ∂K

and ξ + tλ ∈ K

∀t ∈ (t− , t+ )

0 ∩ ∂K. We moreover have that and hence ξ± ∈ K ξ=

−t− t+ ξ+ + ξ− t+ − t− t+ − t−

0 ∩ ∂K. with ξ± ∈ K

(12)

Furthermore, since ξ ∈ ξ0 + α ⊗ εH, we can find γ ∈ εH such that ξ = ξ0 + α ⊗ γ . δ (γ ) ⊂ εH, for every sufficiently small δ > 0. MoreThe set H being open we have that B over, since for every δ > 0, we have 0 ∈ δH = int co{±δβ, δβ3 , . . . , δβk }

Nonconvex problems of the calculus of variations and differential inclusions

91

and since for every sufficiently small δ > 0, we have   ±δβ ∈ co ±(t+ − t− )β ⊂ co ±(t+ − t− )β, δβ3 , . . . , δβk ,

we get that

 0 ∈ δH = int co{±δβ, δβ3 , . . . , δβk } ⊂ int co ±(t+ − t− )β, δβ3 , . . . , δβk .

We are therefore in a position to apply Lemma 48 to a = α,

b = (t+ − t− )β,

bj = δβj

for j = 3, . . . , k, t =

t+ α ⊗ (t+ − t− )β = ξ + (1 − t)a ⊗ b, t+ − t − t− B = ξ− = ξ + α ⊗ (t+ − t− )β = ξ − ta ⊗ b, t+ − t−

−t− , t+ − t−

A = ξ+ = ξ +

 RN ), open sets Ω+ , Ω− ⊂ Ω, such that and find uδ ∈ Affpiec (Ω;   meas(Ω+ ∪ Ω− ) − meas Ω   δ,     u (x) = u (x), x ∈ ∂Ω, and u (x) − u (x)  δ, x ∈ Ω, δ ξ δ ξ  Duδ (x) = ξ± a.e. in Ω± ,    Duδ (x) ∈ ξ + {t+ α ⊗ β, t− α ⊗ β, α ⊗ δβ3 , . . . , α ⊗ δβk } a.e. in Ω.

(13)

0 ∩ ∂K and Since ξ± ∈ K

  = ξ0 + α ⊗ γ + δ H  ξ + α ⊗ δβj ∈ ξ + α ⊗ δ H ⊂ ξ0 + α ⊗ εH ⊂ K0

for j = 3, . . . , k,

we deduce, by choosing δ = 1/ν as ν → ∞, from (13), the relaxation property (12). This achieves the proof of Step 2 and thus of the corollary.  We finally want to point out that, as a particular case of Corollary 47, we find the existence theorem (Theorem 3.1) proved by Dacorogna and Marcellini [27]. We have used the following result due to Müller and Sychev [63] and which is a refinement of a classical result. L EMMA 48 (Approximation lemma). Let Ω ⊂ Rn be a bounded open set. Let t ∈ [0, 1] and A, B ∈ RN ×n such that A−B =a⊗b

92

B. Dacorogna

with a ∈ RN and b ∈ Rn . Let b3 , . . . , bk ∈ Rn , k
n, such that 0 ∈ int co{b, −b, b3 , . . . , bk }. Let ϕ be an affine map such that Dϕ(x) = ξ0 = tA + (1 − t)B,

 x∈Ω

(i.e., A = ξ0 + (1 − t)a ⊗ b and B = ξ0 − ta ⊗ b). Then, for every ε > 0, there exists a piecewise affine map u and there exist disjoint open sets ΩA , ΩB ⊂ Ω, such that    | meas ΩA − t meas Ω|, meas ΩB − (1 − t) meas Ω   ε,        x ∈ ∂Ω, and u(x) − ϕ(x)  ε, x ∈ Ω,  u(x) = ϕ(x),
A in ΩA ,  Du(x) =   B in ΩB ,     a.e. in Ω. Du(x) ∈ ξ0 + (1 − t)a ⊗ b, −ta ⊗ b, a ⊗ b3 , . . . , a ⊗ bk 5.3. Necessary conditions Recall that we are considering the minimization problem






 inf I (u) = f Du(x) dx: u ∈ uξ0 + W01,∞ Ω; RN ,

(P)

Ω

where Ω is a bounded open set of Rn , uξ0 is affine, i.e., Duξ0 = ξ0 and f : RN ×n → R is a lower semicontinuous, locally bounded and nonnegative function. In order to avoid the trivial case we will always assume that Qf (ξ0 ) < f (ξ0 ). Most nonexistence results for problem (P) follow by showing that the relaxed problem (QP) has a unique solution, namely uξ0 , which is by hypothesis not a solution of (P). This approach was strongly used in Marcellini [51], Dacorogna and Marcellini [27] and Dacorogna and Pisante [34]; we will follow here [34]. We should point out that we will give an example (see Proposition 78 in Section 7.5) related to minimal surfaces, where nonexistence occurs, while the relaxed problem has infinitely many solutions, none of them being a solution of (P). The right notion in order to have uniqueness of the relaxed problem is the following definition. D EFINITION 49. A quasiconvex function f : RN ×n → R is said to be strictly quasiconvex at ξ0 ∈ RN ×n , if for some bounded domain U ⊂ Rn the following equality holds 

 f ξ0 + Dϕ(x) dx = f (ξ0 ) meas(U ) U

for some ϕ ∈ W01,∞ (U ; RN ), then necessarily ϕ ≡ 0.

Nonconvex problems of the calculus of variations and differential inclusions

93

We should observe that as in Remark 5(v) the notion of strict quasiconvexity is independent of the choice of the domain U , more precisely we have the following proposition. P ROPOSITION 50. If a function f : RN ×n → R is strictly quasiconvex at ξ0 ∈ RN ×n for one bounded domain U ⊂ Rn it is so for any such domain. P ROOF. Let V ⊂ Rn be a bounded domain and ψ ∈ W01,∞ (V ; RN ) be such that 

 f ξ0 + Dψ(x) dx = f (ξ0 ) meas(V )

(14)

V

and let us conclude that we necessarily have ψ ≡ 0. Choose first a > 0 sufficiently large so that V ⊂ Qa = (−a, a)n and then define v(x) =




ψ(x) 0

if x ∈ V , if x ∈ Qa − V ,

so that v ∈ W01,∞ (Qa ; RN ). Let then x0 ∈ U and choose ν sufficiently large so that  1 a a n x0 + Qa = x0 + − , ⊂ U. ν ν ν Define next ϕ(x) =

1  ν v ν(x − x0 ) 0

Observe that ϕ ∈ W01,∞ (U ; RN ) and 

 f ξ0 + Dϕ(x) dx U

if x ∈ x0 + ν1 Qa , & % if x ∈ U − x0 + ν1 Qa .

(   '

 1 = f (ξ0 ) meas U − x0 + Qa + f ξ0 + Dv ν(x − x0 ) dx 1 ν [x0 + ν Qa ] ' ( 

 meas(Qa ) 1 = f (ξ0 ) meas(U ) − + f ξ0 + Dv(y) dy n n ν ν Qa ( ' meas(Qa ) meas(Qa − V ) + = f (ξ0 ) meas(U ) − νn νn 

 1 f ξ0 + Dψ(y) dy. + n ν V

94

B. Dacorogna

Appealing to (14), we deduce that 

 f ξ0 + Dϕ(x) dx = f (ξ0 ) meas(U ). U

Since f is strictly quasiconvex at ξ0 ∈ RN ×n for the domain U, we deduce that ϕ ≡ 0, which in turn implies that v(y) ≡ 0 for every y ∈ Qa . This finally implies that ψ ≡ 0 as claimed.



We will see further some sufficient conditions that can ensure strict quasiconvexity, but let us start with the elementary following nonexistence theorem. T HEOREM 51. Let f : RN ×n → R be lower semicontinuous, locally bounded and nonnegative, ξ0 ∈ RN ×n with Qf (ξ0 ) < f (ξ0 ) and Qf be strictly quasiconvex at ξ0 . Then the relaxed problem (QP) has a unique solution, namely uξ0 , while (P) has no solution. P ROOF. The fact that (QP) has only one solution follows by definition of the strict quasiconvexity of Qf and Proposition 50. Assume for the sake of contradiction that (P) has a solution u¯ ∈ uξ0 + W01,∞ (Ω; RN ). We should have from Theorem 41 that (writing u(x) ¯ = ξ0 x + ϕ(x))



 f ξ0 + Dϕ(x) = Qf ξ0 + Dϕ(x) a.e. x ∈ Ω, 

 Qf ξ0 + Dϕ(x) dx = Qf (ξ0 ) meas Ω. Ω

Since Qf is strictly quasiconvex at ξ0 , we deduce from the last identity that ϕ ≡ 0. Hence we have, from the first identity, that Qf (ξ0 ) = f (ξ0 ), which is in contradiction with the hypothesis.  We now want to give some criteria that can ensure the strict quasiconvexity of a given function. The first one has been introduced by Dacorogna and Marcellini [27]. D EFINITION 52. A convex function f : RN ×n → R is said to be strictly convex at ξ0 ∈ RN ×n in at least N directions if there exists α = (α i )1iN ∈ RN ×n , α i = 0, for every i = 1, . . . , N , such that: if for some η ∈ RN ×n the identity  1 1 1 f (ξ0 + η) + f (ξ0 ) = f ξ0 + η 2 2 2

holds, then necessarily ! i i" α ; η = 0,

i = 1, . . . , N.

Nonconvex problems of the calculus of variations and differential inclusions

95

In order to understand better the generalization of this notion to polyconvex functions (cf. Proposition 58), it might be enlightening to state the definition in the following way. P ROPOSITION 53. Let f : RN ×n → R be a convex function and, for ξ ∈ RN ×n , denote by ∂f (ξ ) the subdifferential of f at ξ . The two following conditions are then equivalent: (i) f is strictly convex at ξ0 ∈ RN ×n in at least N directions, (ii) there exists α = (α i )1iN ∈ RN ×n with α i = 0 for every i = 1, . . . , N, so that whenever f (ξ0 + η) − f (ξ0 ) − λ; η = 0 for some η ∈ RN ×n and for some λ ∈ ∂f (ξ0 ), then ! i i" α ; η = 0, i = 1, . . . , N. P ROOF.

Step 1. We start with a preliminary observation that if  1 1 1 f (ξ0 + η) + f (ξ0 ) = f ξ0 + η 2 2 2

(15)

then, for every t ∈ [0, 1], we have tf (ξ0 + η) + (1 − t)f (ξ0 ) = f (ξ0 + tη).

(16)

Let us show this under the assumption that t > 1/2 (the case t < 1/2 is handled similarly). We can therefore find α ∈ (0, 1) such that 1 = αt + (1 − α)0 = αt. 2 From the convexity of f and by hypothesis, we obtain  1 1 1 f (ξ0 + η) + f (ξ0 ) = f ξ0 + η  αf (ξ0 + tη) + (1 − α)f (ξ0 ). 2 2 2 Assume, for the sake of contradiction, that f (ξ0 + tη) < tf (ξ0 + η) + (1 − t)f (ξ0 ). Combine then this inequality with the previous one to get % & 1 1 f (ξ0 + η) + f (ξ0 ) < α tf (ξ0 + η) + (1 − t)f (ξ0 ) + (1 − α)f (ξ0 ) 2 2 1 1 = f (ξ0 + η) + f (ξ0 ) 2 2

96

B. Dacorogna

which is clearly a contradiction. Therefore the convexity of f and the above contradiction implies (16). This also implies that f ′ (ξ0 , η) := lim

t→0+

f (ξ0 + tη) − f (ξ0 ) = f (ξ0 + η) − f (ξ0 ). t

Applying Theorem 23.4 in Rockafellar [71], combined with the fact that ∂f (ξ0 ) is nonempty and compact, we get that there exists λ ∈ ∂f (ξ0 ) so that f (ξ0 + η) − f (ξ0 ) = λ; η and hence f (ξ0 + tη) − f (ξ0 ) − tλ; η = 0

∀t ∈ [0, 1].

(17)

We have therefore proved that (15) implies (17). Since the converse is obviously true, we conclude that they are equivalent. Step 2. Let us show the equivalence of the two conditions. (i) ⇒ (ii). We first observe that for any µ ∈ RN ×n we have  1 1 1 f (ξ0 + η) + f (ξ0 ) − f ξ0 + η 2 2 2 ( '  & 1% 1 1 = f (ξ0 + η) − f (ξ0 ) − µ; η − f ξ0 + η − f (ξ0 ) − µ; η . 2 2 2 (18) Assume that, for λ ∈ ∂f (ξ0 ), we have f (ξ0 + η) − f (ξ0 ) − λ; η = 0. From (18) applied to µ = λ, from the definition of ∂f (ξ0 ) and from the convexity of f, we have  1 1 1 0  f (ξ0 + η) + f (ξ0 ) − f ξ0 + η 2 2 2 '  ( 1 1 = − f ξ0 + η − f (ξ0 ) − λ; η  0. 2 2 Using the above identity, we then are in the framework of (i) and we deduce that α i ; ηi  = 0, i = 1, . . . , N, and thus (ii). (ii) ⇒ (i). Assume now that we have (15), namely  1 1 1 f (ξ0 + η) + f (ξ0 ) − f ξ0 + η = 0, 2 2 2 which, by Step 1, implies that there exists λ ∈ ∂f (ξ0 ) so that f (ξ0 + tη) − f (ξ0 ) − tλ; η = 0 ∀t ∈ [0, 1].

Nonconvex problems of the calculus of variations and differential inclusions

97

We are therefore, choosing t = 1, in the framework of (ii) and we get α i ; ηi  = 0, i = 1, . . . , N, as wished.  Of course any strictly convex function is strictly convex in at least N directions, but the above condition is much weaker. For example, in the scalar case, N = 1, it is enough that the function is not affine in a neighborhood of ξ0 , to guarantee the condition (see Corollary 55). We now have the following result established by Dacorogna and Marcellini [27]. P ROPOSITION 54. If a convex function f : RN ×n → R is strictly convex at ξ0 ∈ RN ×n in at least N directions, then it is strictly quasiconvex at ξ0 . Theorem 51, combined with the above proposition, gives immediately a sharp result for the scalar case, namely the following corollary. C OROLLARY 55. Let f : Rn → R be lower semicontinuous, locally bounded and nonnegative, ξ0 ∈ Rn with Cf (ξ0 ) < f (ξ0 ) and Cf not affine in the neighborhood of ξ0 . Then (P) has no solution. R EMARK 56. In the scalar case this result has been obtained by several authors, in particular, by Cellina [16], Friesecke [40] and Dacorogna and Marcellini [27]. It also gives (cf. Theorem 66), combined with the result of the preceding section, that, provided some appropriate boundedness is assumed, a necessary and sufficient condition for existence of minima for (P) is that f be affine on the connected component of {ξ : Cf (ξ ) < f (ξ )} that contains ξ0 . Before proceeding with the proof of Proposition 54 we need the following elementary lemma. L EMMA 57. Let Ω be a bounded open set of Rn and ϕ ∈ W01,∞ (Ω; RN ) be such that " ! i α ; Dϕ i (x) = 0 a.e. x ∈ Ω, i = 1, . . . , N,

for some α i = 0, i = 1, . . . , N , then ϕ ≡ 0.

P ROOF. Working component by component we can assume that N = 1 and therefore we will drop the indices. So let ϕ ∈ W01,∞ (Ω) satisfy for some α ∈ Rn , α = 0, ! " α; Dϕ(x) = 0 a.e. x ∈ Ω.

We then choose α2 , . . . , αn ∈ Rn so that {α, α2 , . . . , αn } generate a basis of Rn . Let a > 0 and for m an integer, m Qm a = (−a, a) .

98

B. Dacorogna

Let x ∈ Ω and let a and t be sufficiently small so that x + τ α + τ2 α2 + · · · + τn αn ∈ Ω

for every τ ∈ (0, t) and (τ2 , . . . , τn ) ∈ Qn−1 a .

Observe then that if ϕ ∈ C01 (Ω), then 

Qn−1 a

= =

% & ϕ(x + tα + τ2 α2 + · · · + τn αn ) − ϕ(x + τ2 α2 + · · · + τn αn ) dτ2 · · · dτn



Qn−1 a



Qn−1 a



0



0

t

& d% ϕ(x + τ α + τ2 α2 + · · · + τn αn ) dτ dτ2 · · · dτn dτ

t!

" Dϕ(x + τ α + τ2 α2 + · · · + τn αn ); α dτ dτ2 · · · dτn .

By a standard regularization procedure the above identity also holds for any ϕ ∈ W01,∞ (Ω). Since α; Dϕ = 0, we deduce that 

Qn−1 a

% ϕ(x + tα + τ2 α2 + · · · + τn αn )

& − ϕ(x + τ2 α2 + · · · + τn αn ) dτ2 · · · dτn = 0.

Since ϕ is continuous, we deduce, by dividing by the measure of Qn−1 and letting a → 0, a that, for every t sufficiently small so that x + tα ∈ Ω, ϕ(x + tα) = ϕ(x). Choosing t so that x + τα ∈ Ω

∀τ ∈ [0, t) and x + tα ∈ ∂Ω,

we obtain the claim, namely ϕ(x) = 0 ∀x ∈ Ω.



P ROOF OF P ROPOSITION 54. Assume that, for a certain bounded domain U ⊂ Rn and for some ϕ ∈ W01,∞ (U ; RN ), we have 

U

 f ξ0 + Dϕ(x) dx = f (ξ0 ) meas(U )

and let us show that ϕ ≡ 0.

Nonconvex problems of the calculus of variations and differential inclusions

99

Since f is convex and the above identity holds, we find  '

(  1 1 f (ξ0 ) meas(U ) = f (ξ0 ) + f ξ0 + Dϕ(x) dx 2 U 2   1 f ξ0 + Dϕ(x) dx
2 U
f (ξ0 ) meas(U ),

which implies that  ' U

(   1 1 1 f (ξ0 ) + f ξ0 + Dϕ(x) − f ξ0 + Dϕ(x) dx = 0. 2 2 2

The convexity of f implies then that, for almost every x in U , we have   1 1 1 f (ξ0 ) + f ξ0 + Dϕ(x) − f ξ0 + Dϕ(x) = 0. 2 2 2 The strict convexity in at least N directions leads to ! i " α ; Dϕ i (x) = 0 a.e. x ∈ Ω, i = 1, . . . , N.

Lemma 57 gives the claim.



We will now generalize Proposition 54. Since the notations in the next result are involved, we will first write the proposition when N = n = 2. P ROPOSITION 58. Let f : RN ×n → R be polyconvex, ξ0 ∈ RN ×n and λ = λ(ξ0 ) ∈ Rτ (N,n) so that ! " f (ξ0 + η) − f (ξ0 ) − λ; T (ξ0 + η) − T (ξ0 )
0 for every η ∈ RN ×n .

(i) Let N = n = 2 and assume that there exist α 1,1 , α 1,2 , α 2,2 ∈ R2 , α 1,1 = 0, α 2,2 = 0, β ∈ R, so that if for some η ∈ R2×2 the following equality holds ! " f (ξ0 + η) − f (ξ0 ) − λ; T (ξ0 + η) − T (ξ0 ) = 0,

then necessarily

! 2,2 2 " ! " ! " α ; η = 0 and α 1,1 ; η1 + α 1,2 ; η2 + β det η = 0.

Then f is strictly quasiconvex at ξ0 .

100

B. Dacorogna

(ii) Let N, n
2 and assume that there exist, for every ν = 1, . . . , N, α ν,ν , α ν,ν+1 , . . . , α ν,N ∈ Rn , n

β ν,s ∈ R(s ) ,

α ν,ν = 0,

2  s  n ∧ (N − ν + 1),

so that if for some η ∈ RN ×n the following equality holds

! " f (ξ0 + η) − f (ξ0 ) − λ; T (ξ0 + η) − T (ξ0 ) = 0,

then necessarily

−ν+1) N  ! ν,s s " n∧(N ! ν,s

" α ;η + β ; adjs ην , . . . , ηN = 0, s=ν

s=2

ν = 1, . . . , N.

Then f is strictly quasiconvex at ξ0 . R EMARK 59. (i) The existence of a λ as in the hypotheses of the proposition is automatically guaranteed by the polyconvexity of f (see (1) in Section 2, it corresponds in the case of a convex function to an element of ∂f (ξ0 )). (ii) We have adopted the convention that if l > k > 0 are integers, then k  l

= 0.

E XAMPLE 60. Let N = n = 2 and consider the function 2

2 f (η) = η22 + η11 + det η .

This function is trivially polyconvex and according to the proposition it is also strictly quasiconvex at ξ0 = 0 (choose λ = 0 ∈ R5 , α 2,2 = (0, 1), α 1,2 = (0, 0), α 1,1 = (1, 0), β = 1). P ROOF OF P ROPOSITION 58. We will prove the proposition only in the case N = n = 2, the general case being handled similarly. Assume that, for a certain bounded domain U ⊂ R2 and for some ϕ ∈ W01,∞ (U ; R2 ), we have 

 f ξ0 + Dϕ(x) dx = f (ξ0 ) meas(U ) U

and let us prove that ϕ ≡ 0. This is equivalent, for every µ ∈ Rτ (2,2) , to '

U

(

 !

 " f ξ0 + Dϕ(x) − f (ξ0 ) − µ; T ξ0 + Dϕ(x) − T (ξ0 ) dx = 0.

Nonconvex problems of the calculus of variations and differential inclusions

101

Choosing µ = λ (λ as in the statement of the proposition) in the previous equation and using the polyconvexity of the function f , we get

 !

 " f ξ0 + Dϕ(x) − f (ξ0 ) − λ; T ξ0 + Dϕ(x) − T (ξ0 ) = 0 a.e. x ∈ Ω.

We hence infer that, for almost every x ∈ Ω, we have

! 2,2 " ! " ! " α ; Dϕ 2 = 0 and α 1,1 ; Dϕ 1 + α 1,2 ; Dϕ 2 + β det Dϕ = 0.

Lemma 57, applied to the first equation, implies that ϕ 2 ≡ 0. Using this result in the second equation we get ! 1,1 " α ; Dϕ 1 = 0

and hence, appealing once more to the lemma, we have the claim, namely ϕ 1 ≡ 0.



Summarizing the results of Theorem 51, Propositions 54 and 58, we get the following corollary. C OROLLARY 61. Let f : RN ×n → R be lower semicontinuous, locally bounded and nonnegative, ξ0 ∈ RN ×n with Qf (ξ0 ) < f (ξ0 ). If either one of the two following conditions hold (i) Qf (ξ0 ) = Cf (ξ0 ) and Cf is strictly convex at ξ0 in at least N directions, (ii) Qf (ξ0 ) = Pf (ξ0 ) and Pf is strictly polyconvex at ξ0 (in the sense of Proposition 58), then (QP) has a unique solution, namely uξ0 , while (P) has no solution. P ROOF. The proof is almost identical under both hypotheses and so we will establish the corollary only in the first case. The result will follow from Theorem 51 if we can show that Qf is strictly convex at ξ0 . So assume that 

Ω

 Qf ξ0 + Dϕ(x) dx = Qf (ξ0 ) meas Ω

for some ϕ ∈ W01,∞ (Ω; RN ) and let us prove that ϕ ≡ 0. Using the Jensen inequality combined with the hypothesis Qf (ξ0 ) = Cf (ξ0 ) and the fact that Qf
Cf , we find that the above identity implies 

Ω

 Cf ξ0 + Dϕ(x) dx = Cf (ξ0 ) meas Ω.

The hypotheses on Cf and Proposition 54 imply that ϕ ≡ 0, as wished.



102

B. Dacorogna

We now conclude this section with a different necessary condition that is based on the Carathéodory theorem. Recall first that for any integer s, we let   s  λi = 1 . Λs = λ = (λ1 , . . . , λs ): λi
0 and i=1

T HEOREM 62. If (P) has a solution u¯ ∈ uξ0 + W01,∞ (Ω; RN ), then there exist µ ∈ ΛN n+1 and ξν ∈ RN ×n , |ξν |  u ¯ W 1,∞ , 1  ν  N n + 1, such that Qf (ξ0 )


N n+1

µν f (ξν )

ν=1

and ξ0 =

N n+1

µ ν ξν .

ν=1

Moreover, if either n = 1 or N = 1, the inequality becomes an equality, namely N n+1

Cf (ξ0 ) =

µν f (ξν )

ν=1

and ξ0 =

N n+1

µ ν ξν .

ν=1

R EMARK 63. The theorem is just a curiosity in the vectorial case n, N > 1. However in the scalar case, n > N = 1, under some extra hypotheses (cf. Theorem 66), one of them being ξ0 ∈ int co{ξ1 , . . . , ξn+1 }, it turns out that the necessary condition is also sufficient. But it is in the case N
n = 1 that it is particularly interesting since then this condition is also sufficient, cf. Theorem 64. P ROOF OF T HEOREM 62. We decompose the proof into three steps. Step 1. Let u¯ ∈ uξ0 + W01,∞ (Ω; RN ) be a solution of (P). It should therefore satisfy 1 meas Ω



Ω

 f D u(x) ¯ dx = inf(P) = inf(QP) = Qf (ξ0 ).

(19)

Let r = u ¯ W 1,∞ and use the fact that f is locally bounded to find R = R(r) so that

 0  f D u(x) ¯  R a.e. x ∈ Ω.

Denote by

 Kr = (ξ, y) ∈ RN ×n × R: |ξ |  r and |y|  R ,  epi f = (ξ, y) ∈ RN ×n × R: f (ξ )  y , E = epi f ∩ Kr .

Nonconvex problems of the calculus of variations and differential inclusions

103

Note that since f is lower semicontinuous then epi f is closed and hence E is compact. Therefore its convex hull co E is also compact. Observe that, for almost every x ∈ Ω, we have

 D u(x), ¯ f D u(x) ¯ ∈E

and thus by the Jensen inequality and (19) we deduce that

 ξ0 , Qf (ξ0 ) =

1 meas Ω



Ω

 D u(x), ¯ f D u(x) ¯ dx ∈ co E.

Appealing to the Carathéodory theorem we can find λ ∈ ΛN n+2 , (ξi , yi ) ∈ E, 1  i  N n + 2 (in particular, f (ξi )  yi ), such that Qf (ξ0 ) =

N n+2

λi yi


N n+2 i=1

i=1

λi f (ξi ) and ξ0 =

N n+2

λi ξi .

i=1

(Note, in passing, that if f is continuous, we can replace in the above argument epi f by  graph f = (x, y) ∈ RN ×n × R: f (x) = y

obtaining therefore equality instead of inequality in the above statement.) Step 2. To obtain the theorem it therefore remains to show that one can take only (N n + 1) elements. This is a classical procedure in convex analysis. The result is equivalent to showing that there exist µi , 1  i  N n + 2, such that 

µi
0, N n+2 i=1

N n+2

µi = 1, at least one of the µi = 0, N n+2 N n+2 µ i ξi , ξ0 = i=1 µi f (ξi )  i=1 λi f (ξi ), i=1

(20)

meaning in fact that µ ∈ ΛN n+1 as wished. Assume that λi > 0, 1  i  N n + 2, otherwise nothing is to be proved. Observe first that ξ0 ∈ co{ξ1 , . . . , ξN n+2 } ⊂ RN ×n . Thus it follows from the Carathéodory theorem that there exist ν ∈ ΛN n+2 , with at least one of the νi = 0 (i.e., ν ∈ ΛN n+1 ), such that ξ0 =

N n+2

νi ξi .

i=1

Assume, without loss of generality, that N n+2 i=1

νi f (ξi ) >

N n+2 i=1

λi f (ξi );

(21)

104

B. Dacorogna

otherwise choosing µi = νi we would have immediately (20). Let  J = i ∈ {1, . . . , N n + 2}: λi − νi < 0 .

Observe that J = ∅, since otherwise λi
νi 
0 for every i and since at least one of the νi = 0, we would have a contradiction with νi = λi = 1 and λi > 0 for every i. We then define


λi γ = min . i∈J νi − λi We clearly have that γ > 0. Finally let µi = λi + γ (λi − νi ),

1  i  N n + 2.

We immediately get that N n+2

µi
0,

i=1

µi = 1,

at least one of the µi = 0.

(22)

From (21) we obtain N n+2 i=1

µi f (ξi ) = 

N n+2

λi f (ξi ) + γ

N n+2

λi f (ξi ).

i=1

N n+2  i=1

λi f (ξi ) −

N n+2 i=1



νi f (ξi )

i=1

The combination of the above with (22) (assuming for the sake of notation that µN n+2 = 0) gives immediately Qf (ξ0 )


N n+1 i=1

µi f (ξi )

and ξ0 =

N n+1

µ i ξi .

i=1

Step 3. The result for the scalar case follows from the fact that Qf (ξ0 ) = Cf (ξ0 ) and from Theorem 6.  6. The scalar case We now see how to apply the above abstract considerations to the case where either n = 1 or N = 1. We recall that




  1,∞ N inf I (u) = . (P) f Du(x) dx: u ∈ uξ0 + W0 Ω; R Ω

Nonconvex problems of the calculus of variations and differential inclusions

105

We will first treat the more elementary case where n = 1 and then the case N = 1. 6.1. The case of single integrals In this very elementary case we can get much simpler and sharper results. T HEOREM 64. Let f : RN → R be nonnegative, locally bounded and lower semicontinuous. Let a < b, α, β ∈ RN , N
1, and
 inf I (u) =

b

a



 f u′ (x) dx: u ∈ X ,

(P)

where 

 X = u ∈ W 1,∞ (a, b); RN : u(a) = α, u(b) = β .

The two following statements are then equivalent: (i) problem (P) has a minimizer, N +1 λν = 1, γν ∈ RN , 1  ν  N + 1, such that (ii) there exist λν
0 with ν=1 

β −α Cf b−a



=

N +1 

λν f (γν )

and

ν=1

N +1 β −α  = λ ν γν , b−a

(23)

ν=1

where Cf = sup{g  f : g convex}. Furthermore, if (23) is satisfied and if )

Ip = a + (b − a)

p−1  ν=1

λν , a + (b − a)

p  ν=1

+

λν ,

1  p  N + 1,

then u(x) ¯ = γp (x − a) + (b − a)

p  ν=1

λν (γν − γp ) + α,

x ∈ Ip , 1  p  N + 1,

is a solution of (P). R EMARK 65. (i) The sufficiency of (23) is implicitly or explicitly proved in the papers mentioned in the bibliography. The necessity is less known but is also implicit in the literature. The theorem as stated can be found in Dacorogna [25].

106

B. Dacorogna

(ii) Recall that by the Carathéodory theorem (cf. Theorem 6) we always have 

β −α Cf b−a



N +1  N +1   β −α = inf λν f (γν ): λ ν γν = . b−a ν=1

(24)

ν=1

Therefore (23) states that a necessary and sufficient condition for existence of solutions is that the infimum in (24) be attained. Note also that if f is convex or f coercive (in the sense that f (ξ )
a|ξ |p + b with p > 1, a > 0) then the infimum in (24) is always attained. (iii) Therefore if f (x, u, ξ ) = f (ξ ), counterexamples to existence must be nonconvex and noncoercive; cf. Example 1, where
 inf I (u) =

1

0

′ 2 e−(u (x)) dx: u ∈ W01,∞ (0, 1) ,

(P)

2

i.e., f (ξ ) = e−ξ , then Cf (ξ ) ≡ 0 and therefore, by the relaxation theorem, inf(P) = inf(QP) = 0. However it is obvious that I (u) = 0 for every u ∈ W01,∞ (0, 1) and hence the infimum of (P) is not attained. (iv) A similar proof to that of Theorem 64 (see, for example, Marcellini [50]) shows that a sufficient condition to ensure existence of minima to


 b

 ′ f x, u (x) dx: u ∈ X (P) inf I (u) = a

is (23), where λν and γν are then measurable functions. Of course if f depends explicitly on u, the example of Bolza (cf. Example 2) shows that the theorem is then false. P ROOF OF T HEOREM 64. It is easy to see that we can reduce our study to the case where a = 0,

b = 1 and α = 0.

Sufficient condition. The sufficiency part is elementary. Let
 inf I(u) =

0

1



 Cf u′ (x) dx: u ∈ X ,

where now 

 X = u ∈ W 1,∞ (0, 1); RN : u(0) = 0, u(1) = β .

Then u(x) ˜ = βx is trivially a solution of (QP) and therefore inf(QP) = Cf (β).

(QP)

Nonconvex problems of the calculus of variations and differential inclusions

107

Let now u¯ be as in the statement of the theorem. Observe first that u¯ ∈ W 1,∞ ((0, 1); RN ) and u(0) ¯ = 0, u(1) ¯ = β. We now compute I(u) ¯ = =



1

0

N +1 

 f u¯ ′ (x) dx =

N +1  p=1



p=1 Ip

N +1 

 f (γp ) meas Ip f u¯ ′ (x) dx = p=1

λp f (γp ) = Cf (β) = inf(QP)  inf(P).

Necessary condition. This has already been proved in Theorem 62.



6.2. The case of multiple integrals We now discuss the case n > N = 1. This is of course a more difficult case than the preceding one and no such simple result as Theorem 64 is available. However, we immediately have from Sections 5.2 and 5.3 (Theorem 29 and Corollary 55) the theorem stated below. For some historical comments on this theorem, see the remark following Corollary 55. But let us first recall the problem and the notation. We have




 1,∞ inf I (u) = f Du(x) dx: u ∈ uξ0 + W0 (Ω) ,

(P)

Ω

where Ω is a bounded open set of Rn , uξ0 is affine, i.e., Duξ0 = ξ0 and f : Rn → R is a lower semicontinuous, locally bounded and nonnegative function. Let Cf = sup{g  f : g convex}. In order to avoid the trivial situation we assume that Cf (ξ0 ) < f (ξ0 ). We next set  K = ξ ∈ Rn : Cf (ξ ) < f (ξ )

and we assume that it is connected, otherwise we replace it by its connected component that contains ξ0 . T HEOREM 66. Necessary condition. If (P) has a minimizer, then Cf is affine in a neighborhood of ξ0 . Sufficient condition. If there exists E ⊂ ∂K such that ξ0 ∈ int co E and Cf |E∪{ξ0 } is affine, then (P) has a solution.

108

B. Dacorogna

R EMARK 67. (i) By Cf |E∪{ξ0 } affine we mean that there exist α ∈ Rn , β ∈ R such that Cf (ξ ) = α; ξ  + β

for every ξ ∈ E ∪ {ξ0 }.

Usually one proves that Cf is affine on the whole of co E. (ii) The theorem applies, of course, to the case where E = ∂K and Cf is affine on the whole of K (since K is open and ξ0 ∈ K ⊂ int co K). However in many simple examples such as the one given below, it is not realistic to assume that E = ∂K. P ROOF OF T HEOREM 66. The necessary part is just Corollary 55. We therefore discuss only the sufficient part. We use Theorem 29 to find u¯ ∈ uξ0 + W01,∞ (Ω) such that D u(x) ¯ ∈ E ⊂ ∂K

a.e. x ∈ Ω

and hence



 f D u(x) ¯ = Cf D u(x) ¯

a.e. x ∈ Ω.

Then use the fact that Cf |E∪{ξ0 } is affine to deduce that 

Ω

 Cf D u(x) ¯ dx = Cf (ξ0 ) meas Ω.

The conclusion then follows from Theorem 41.



We now would like to give two simple examples. The first one generalizes Example 3. E XAMPLE 68. Let N = 1, n = 2, Ω = (0, 1)2 , u0 ≡ 0, a
0 and

2 2 f (ξ ) = ξ12 − 1 + ξ22 − a 2 .

We find that

where

&2 % &2 % Cf (ξ ) = ξ12 − 1 + + ξ22 − a 2 + , [x]+ =




x 0

if x
0, if x < 0.

We therefore have that  K = ξ ∈ R2 : ξ12 < 1 or ξ22 < a 2

and note that it is unbounded and that Cf is not affine on the whole of K. Let us discuss the two different cases.

Nonconvex problems of the calculus of variations and differential inclusions

109

Case 1: a = 0. This corresponds to Example 3. Then clearly Cf is not affine in the neighborhood of ξ0 = 0, since it is strictly convex in the direction e2 = (0, 1). Hence (P) has no solution. Case 2: a > 0. We let  E = ξ ∈ R2 : |ξ1 | = 1 and |ξ2 | = a ⊂ ∂K.

Note that ξ0 = 0 ∈ int co E and Cf |co E ≡ 0 is affine. Therefore the theorem applies and we obtain that (P) has a solution. E XAMPLE 69. We conclude with the following example (cf. Marcellini [51] and Dacorogna and Marcellini [27]). Let n
2 and

 f (Du) = g |Du| ,

where g : R → R is lower semicontinuous, locally bounded and nonnegative with  g(0) = inf g(t) : t
0 .

It is easy to see that Cf = Cg. Let  S = t
0: Cg(t) < g(t) ,   K = ξ ∈ Rn : Cf (ξ ) < f (ξ ) = ξ ∈ Rn : |ξ | ∈ S .

Assume that ξ0 ∈ K and that S is connected, otherwise replace it by its connected component containing |ξ0 |. We then have to consider two cases. Case 1: Cg is strictly increasing at |ξ0 |. Then clearly Cf is not affine in any neighborhood of ξ0 and hence (P) has no solution. Case 2 : Cg is constant on S. Assume that S is bounded, this can be guaranteed if, for example, lim

t→+∞

g(t) = +∞. t

So let |ξ0 | ∈ S = (α, β) and choose in the sufficient part of the theorem  E = ξ ∈ Rn : |ξ | = β

and apply the theorem to find a minimizer for (P).

110

B. Dacorogna

7. The vectorial case We now consider several examples of the form studied in the previous sections, namely






 inf I (u) = f Du(x) dx: u ∈ uξ0 + W01,∞ Ω; RN ,

(P)

Ω

where Ω is a bounded open set of Rn , uξ0 is affine, i.e., Duξ0 = ξ0 and f : RN ×n → R is a lower semicontinuous, locally bounded and nonnegative function. 1. We consider in Section 7.1 the case where N = n and 

f (ξ ) = g λ2 (ξ ), . . . , λn−1 (ξ ), det ξ ,

where 0  λ1 (ξ )  · · ·  λn (ξ ) are the singular values of ξ ∈ Rn×n . 2. In Section 7.2 we deal with the case

 f (ξ ) = g Φ(ξ ) ,

where Φ : RN ×n → R is quasiaffine (so in particular, we can have, when N = n, Φ(ξ ) = det ξ , as in the previous case). 3. We next discuss in Section 7.3 the Saint Venant–Kirchhoff energy functional. Up to rescaling, the function under consideration is (here N = n and ν ∈ (0, 1/2) is a parameter) 2  f (ξ ) = ξ ξ ⊤ − I  +

2 ν 2 |ξ | − n 1 − 2ν

or in terms of the singular values, 0  λ1 (ξ )  · · ·  λn (ξ ), of ξ ∈ Rn×n , n 

2 2 f (ξ ) = λi − 1 + i=1

2  n  ν 2 λi − n . 1 − 2ν i=1

4. In Section 7.4 we consider a problem of optimal design where N = n = 2 and f (ξ ) =




1 + |ξ |2 0

if ξ =  0, if ξ = 0.

5. In Section 7.5 we deal with the minimal surface case, namely when N = n + 1 and f (ξ ) = g(adjn ξ ). 6. Finally in Section 7.6 we discuss the problem of potential wells. We recall that, throughout Section 7, the sets O(n) and SO(n) will denote respectively the set of orthogonal and special orthogonal matrices, more presicely,  O(n) = R ∈ Rn×n : RR ⊤ = 1 ,  SO(n) = R ∈ O(n): det R = 1 .

Nonconvex problems of the calculus of variations and differential inclusions

111

7.1. The case of singular values In this section we let N = n and we denote by λ1 (ξ ), . . . , λn (ξ ) the singular values of ξ ∈ Rn×n with 0  λ1 (ξ )  · · ·  λn (ξ ) and by Q the set  Q = x = (x2 , . . . , xn−1 ) ∈ Rn−2 : 0  x2  · · ·  xn−1 ,

which is the natural set where to consider (λ2 (ξ ), . . . , λn−1 (ξ )) for ξ ∈ Rn×n . The functions under consideration are of the form studied in Theorem 10, namely

 f (ξ ) = g λ2 (ξ ), . . . , λn−1 (ξ ), det ξ ,

and we have

Pf (ξ ) = Qf (ξ ) = Rf (ξ ) = Ch(det ξ ), where h : R → R is given by h(s) = infx∈Q g(x, s). We next apply the theory of Section 5.2 to get the following existence result established by Dacorogna, Pisante and Ribeiro [34]. T HEOREM 70. Let

 f (ξ ) = g λ2 (ξ ), . . . , λn−1 (ξ ) + h(det ξ ),

where g : Q → R is nonnegative, continuous and verifies inf g = g(m2 , . . . , mn−1 )

with 0 < m2  · · ·  mn−1

and h : R → R is a nonnegative, lower semicontinuous and locally bounded function such that lim

|t|→+∞

h(t) = +∞. |t|

Then (P) has a solution. P ROOF. We note that, by Theorem 10, Qf (ξ ) = inf g + Ch(det ξ ). Letting  K = ξ ∈ Rn×n : Qf (ξ ) < f (ξ )

we see that

K = L1 ∪ L 2 ,

(25)

112

B. Dacorogna

where  L1 = ξ ∈ Rn×n : Ch(det ξ ) < h(det ξ ) , 

 L2 = ξ ∈ Rn×n : Ch(det ξ ) = h(det ξ ), inf g < g λ2 (ξ ), . . . , λn−1 (ξ ) .

We now prove the result. Clearly, if ξ0 ∈ / K then uξ0 is a solution of (P), so from now on we assume that ξ0 ∈ K. There are three different cases to consider, one of them will be treated with Theorem 42 and the two others with Theorem 41. Case 1: ξ0 ∈ L1 . We first observe that hypothesis (25) allows us to write  * (αj , βj ), S = t ∈ R: Ch(t) < h(t) = j ∈N

Ch being affine in each interval (αj , βj ); thus Qf is quasiaffine on each connected component of L1 and


* (αj , βj ) . L1 = ξ ∈ Rn×n : det ξ ∈ j ∈N

Let (αj , βj ) be an interval as above such that det ξ0 ∈ (αj , βj ). We get the result applying Theorem 42 with   n n # # n×n mi , ν = 2, . . . , n , λi (ξ ) < K0 = ξ ∈ R : det ξ ∈ (αj , βj ), i=ν

i=ν

where mn is chosen sufficiently large so that mn−1  mn , n n # # λi (ξ0 ) < mi , i=ν

i=ν

(26) ν = 2, . . . , n,

n #  max |αj |, |βj | < m2 mi .

(27) (28)

i=2

Clearly K0 ⊂ L1 ⊂ K, moreover, (27) ensures that ξ0 ∈ K0 and (28) ensures the relaxation property of K0 with respect to  0 ∩ ∂K E = ξ ∈ Rn×n : det ξ ∈ {αj , βj }, λν (ξ ) = mν , ν = 2, . . . , n ⊂ K

through Theorems 21 and 28 and the family of sets

 Eδ = ξ ∈ Rn×n : det ξ ∈ {αj + δ, βj − δ}, λi (ξ ) = mi − δ, i = 2, . . . , n

Nonconvex problems of the calculus of variations and differential inclusions

113

(cf. the proof of Theorem 1.1 of Dacorogna and Ribeiro [35] for details). Consequently 0 ∩ ∂K. K0 has the relaxation property with respect to K Case 2: ξ0 ∈ L2 and det ξ0 = 0. We consider in this case the set 

K1 = ξ ∈ R

n×n

: det ξ = det ξ0 ,

n #

λi (ξ )
mn−1 ). It was shown by Dacorogna and Tanteri [37] that K1 has the relaxation property with respect to  E = ξ ∈ Rn×n : det ξ = det ξ0 , λν (ξ ) = mν , ν = 2, . . . , n ,

and moreover, there exists u ∈ uξ0 + W01,∞ (Ω, Rn ) such that Du ∈ E a.e. in Ω. Since Qf = f in E and Qf (ξ0 ) = Qf (Du), we can apply Theorem 41 and get the result. Case 3: ξ0 ∈ L2 and det ξ0 = 0. We here just briefly outline the idea and we refer to Dacorogna, Pisante and Ribeiro [34] for details. Since any matrix ξ ∈ Rn×n can be decomposed in the form RDQ, where R, Q ∈ O(n) and D = diag(λ1 (ξ ), . . . , λn (ξ )) (cf. [45]) we can reduce ourselves to the case of ξ0 = diag(λ1 (ξ0 ), . . . , λn (ξ0 )). In particular, as det ξ0 = 0, we have λ1 (ξ0 ) = 0 and thus the first line of ξ0 equal to zero. Let mn
mn−1 and define   n n # # n×n 1 mi , ν = 2, . . . , n , λi (ξ ) < K1 = ξ ∈ R : ξ = 0, i=ν

i=ν

 E = ξ ∈ Rn×n : ξ 1 = 0, λi (ξ ) = mi , i = 2, . . . , n ,

we get that K1 has the relaxation property with respect to E. If we choose mn sufficiently large such that ξ0 ∈ K1 we can apply Theorem 25 to get the existence of u ∈ uξ0 + W01,∞ (Ω, Rn ) such that Du ∈ E. Finally, as Qf = f in E and Qf (ξ0 ) = Qf (Du), applying Theorem 41, we conclude the proof. 

7.2. The case of quasiaffine functions We next study the minimization problem


 

 1,∞ N , Ω; R g Φ Du(x) dx: u ∈ uξ0 + W0 inf Ω

where Ω is a bounded open set of Rn , Duξ0 = ξ0 and • g : R → R is a lower semicontinuous, locally bounded and nonnegative function, • Φ : RN ×n → R is quasiaffine and nonconstant.

(P)

114

B. Dacorogna

We recall that in particular we can have, when N = n, Φ(ξ ) = det ξ . The relaxed problem is then


  1,∞ N inf , Cg Φ Du(x) dx: u ∈ uξ0 + W0 Ω; R

(QP)

Ω

where Cg is the convex envelope of g (here f (ξ ) = g(Φ(ξ )) and we get Qf = Cg, cf. Theorem 9). The existence result is the following theorem. T HEOREM 71. Let Ω ⊂ Rn be a bounded open set with Lipschitz boundary, g : R → R a nonnegative, lower semicontinuous and locally bounded function such that lim

|t|→+∞

g(t) = +∞ |t|

(29)

and uξ0 (x) = ξ0 x, with ξ0 ∈ RN ×n . Then there exists u¯ ∈ uξ0 + W01,∞ (Ω; RN ) solution of (P). R EMARK 72. This result has first been established by Mascolo and Schianchi [55] and later by Dacorogna and Marcellini [27] for the case of the determinant. The general case is due to Cellina and Zagatti [19] and later to Dacorogna and Ribeiro [35]. Here we see that it can be obtained as a particular case of Theorem 42. P ROOF OF T HEOREM 71. We will here only sketch the proof and we refer for details to Dacorogna and Ribeiro [35]. We first let  S = t ∈ R: Cg(t) < g(t) .

From the hypothesis on g we can write S=

*

(αj , βj )

j ∈N

with Cg affine in each interval (αj , βj ) and thus Qf is quasiaffine on each connected component of K, where  K = ξ ∈ RN ×n : Φ(ξ ) ∈ S .

/ S then uξ0 is a solution of (P). In the other case, Φ(ξ0 ) ∈ (α, β) ⊂ S for some If Φ(ξ0 ) ∈ α and β, and we apply Theorem 42 with    K0 = ξ ∈ RN ×n : Φ(ξ ) ∈ (α, β), ξji  < cji , i = 1, . . . , N, j = 1, . . . , n ,

Nonconvex problems of the calculus of variations and differential inclusions

115

where cji are constants sufficiently large so that ξ0 ∈ K0 and satisfying      inf Φ(ξ ): ξji  = cji > max |α|, |β| .

This condition allows us to obtain the relaxation property of K0 with respect to    0 ∩ ∂K = ξ ∈ RN ×n : Φ(ξ ) ∈ {α, β}, ξ i   ci , i = 1, . . . , N, j = 1, . . . , n . K j j

The relaxation property is obtained using the approximation property (cf. Definition 27 and Theorem 28) considering the sets, here δ > 0 is sufficiently small,  Hδ = ξ ∈ RN ×n : Φ(ξ ) ∈ {α + δ, β − δ},  i ξ   ci − δ, i = 1, . . . , N, j = 1, . . . , n . j j

This concludes the proof of the theorem.



The problem under consideration is sufficiently flexible that we could also proceed as in Dacorogna and Marcellini [27], using Corollary 47. Indeed if DΦ(ξ0 ) = 0 (in the case Φ(ξ ) = det ξ this means that rank ξ0
n − 1), we can apply the corollary, since the connected component of K containing ξ0 is bounded, in the neighborhood of ξ0 , in a direction of rank one. We do not discuss the details of this different approach.

7.3. The Saint Venant–Kirchhoff energy The problem is now of the form


 

inf f Du(x) dx: u ∈ uξ0 + W01,∞ Ω; Rn ,

(P)

Ω

where, upon rescaling, the function under consideration is, ν ∈ (0, 1/2) being a parameter, 2  f (ξ ) = ξ ξ ⊤ − I  +

2 ν 2 |ξ | − n 1 − 2ν

or in terms of the singular values, 0  λ1 (ξ )  · · ·  λn (ξ ), of ξ ∈ Rn×n , n 

2 2 f (ξ ) = λi − 1 + i=1

2  n  ν 2 λi − n . 1 − 2ν i=1

According to Le Dret and Raoult [49] the quasiconvex envelope and the convex envelope coincide, at least when n = 2 or n = 3, i.e., Qf (ξ ) = Cf (ξ ).

116

B. Dacorogna

In the case n = 2, it is given by  f (ξ )   2

2 1 Qf (ξ ) = 1−ν λ2 − 1   0

if ξ ∈ / D1 ∪ D2 , if ξ ∈ D2 ,

if ξ ∈ D1 ,

where

 % &2 % &2 D1 = ξ ∈ R2×2 : (1 − ν) λ1 (ξ ) + ν λ2 (ξ ) < 1 and λ2 (ξ ) < 1  = ξ ∈ R2×2 : λ1 (ξ )  λ2 (ξ ) < 1 ,  % &2 % &2 D2 = ξ ∈ R2×2 : (1 − ν) λ1 (ξ ) + ν λ2 (ξ ) < 1 and λ2 (ξ )
1 .

The existence theorem is the following one.

T HEOREM 73. Let Ω ⊂ R2 , f and ξ0 be as above. (i) If ξ0 ∈ / D2 then (P) has a solution. (ii) If ξ0 ∈ int D2 then (P) has no solution. R EMARK 74. The nonexistence part has been proved by Dacorogna and Marcellini [27]. P ROOF OF T HEOREM 73. (i) The case where ξ0 ∈ / D1 ∪ D2 corresponds to the trivial case, where Qf (ξ0 ) = f (ξ0 ). So we now assume that ξ0 ∈ D1 . Note that Qf is quasiaffine on D1 (in fact Qf (ξ ) ≡ 0). Apply then Theorem 33 (and the remark following it) to get u ∈ uξ0 + W01,∞ (Ω; R2 ) such that λ1 (Du) = λ2 (Du) = 1 a.e. in Ω. This implies that Qf (Du) = f (Du) = Qf (ξ0 ) = 0 and hence the claim follows from Theorem 41. (ii) It was shown in [27], and we do not discuss here the details, that if ξ0 ∈ int D2 then the function Qf is strictly quasiconvex at ξ0 and therefore (P) has no solution.  7.4. An optimal design problem We now consider the case, studied by many authors following the pioneering work of Kohn and Strang [48], where




 1,∞ 2 inf Ω; R , (P) f Du(x) dx: u ∈ uξ0 + W0 Ω

Ω ⊂ R2 is a bounded open set with Lipschitz boundary, Duξ0 = ξ0 and f (ξ ) =




1 + |ξ |2 0

if ξ =  0, if ξ = 0.

Nonconvex problems of the calculus of variations and differential inclusions

117

It was shown by Kohn and Strang [48] that the quasiconvex envelope is then  1 + |ξ |2 if |ξ |2 + 2| det ξ |
1, Qf (ξ ) =

2 1/2 2 |ξ | + 2| det ξ | − 2| det ξ | if |ξ |2 + 2| det ξ | < 1.

The existence of minimizers for problem (P) was then established by Dacorogna and Marcellini in [27] and [31]. Later Dacorogna and Tanteri [37] gave a different proof which is more in the spirit of the present report and we follow here this last approach.

T HEOREM 75. Let Ω ⊂ R2 , f and ξ0 be as above. Then a necessary and sufficient condition for (P) to have a solution is that one of the following conditions hold: (i) ξ0 = 0 or |ξ0 |2 + 2| det ξ0 |
1 (i.e., f (ξ0 ) = Qf (ξ0 )), (ii) det ξ0 = 0. P ROOF. We do not discuss the details and in particular not the necessary part (see [27] for details). So we assume that we are in the nontrivial case det ξ0 = 0

and |ξ0 |2 + 2| det ξ0 | < 1.

(30)

the We just point out how to define the set K0 of Theorem 42. We have (denoting by R2×2 s set of 2 × 2 symmetric matrices)  K = ξ ∈ R2×2 : |ξ |2 + 2| det ξ | < 1 \ {0}  K0 = ξ ∈ Rs2×2 : det ξ > 0 and trace ξ ∈ (0, 1) ,  0 ∩ ∂K = {0} ∪ ξ ∈ R2×2 K s : det ξ
0 and trace ξ = 1  = ξ ∈ R2×2 s : det ξ
0 and trace ξ ∈ {0, 1} .

Since f is invariant under rotations and symmetries and (30) holds, we can assume, without loss of generality, that ξ0 ∈ K0 . Furthermore Qf is quasiaffine on K0 (Qf (ξ ) = 2 trace ξ − 2 det ξ ), while it is not so on K. It remains to prove that K0 has the relaxation property with 0 ∩ ∂K; and this is easily established as in [37]. respect to K  7.5. The minimal surface case Following Dacorogna, Pisante and Ribeiro [34], we now deal with the case where N = n + 1 and f (ξ ) = g(adjn ξ ). The minimization problem is then




  1,∞ n+1 inf g adjn Du(x) dx: u ∈ uξ0 + W0 , Ω; R Ω

(P)

118

B. Dacorogna

where Ω is a bounded open set of Rn , Duξ0 = ξ0 and g : Rn+1 → R is a nonnegative, lower semicontinuous and locally bounded nonconvex function. From Theorem 9 we have Qf (ξ ) = Cg(adjn ξ ). We next set  S = y ∈ Rn+1 : Cg(y) < g(y)

and assume, in order to avoid the trivial situation, that adjn ξ0 ∈ S. We also assume that S is connected, otherwise we replace it by its connected component that contains adjn ξ0 . Observe that   K = ξ ∈ R(n+1)×n : Qf (ξ ) < f (ξ ) = ξ ∈ R(n+1)×n : adjn ξ ∈ S .

T HEOREM 76. If S is bounded, Cg is affine in S and rank ξ0
n − 1, then (P) has a solution. R EMARK 77. The fact that Cg be affine in S is not a necessary condition for existence of minima, as seen in Proposition 78. P ROOF OF T HEOREM 76. The result follows if we choose a convenient rank-one direction λ = α ⊗ β ∈ R(n+1)×n satisfying the hypothesis of Corollary 47. We remark that, since we suppose Cg affine in S, Qf is quasiaffine in LK (ξ0 + α ⊗ Bε , λ) (cf. Notation 44 and Definition 45) independently of the choice of λ. So we only have to prove that K is stably bounded at ξ0 in a direction λ = α ⊗ β. Firstly we observe that we can find (cf. Theorem 3.1.1 in [45]) P ∈ O(n + 1), Q ∈ SO(n) and 0  λ1  · · ·  λn , so that ξ0 = P LQ,

1in+1

where L = (λj δij )1j n ;

in particular when n = 2, we have L=



λ1 0 0

0 λ2 0



.

Since rank ξ0
n − 1 we have that λ2 > 0. We also note that 

 adjn ξ0 = adjn P · adjn L and adjn L =  

0 .. . 0 (−1)n λ

1 · · · λn



 . 

Nonconvex problems of the calculus of variations and differential inclusions

119

Without loss of generality we assume ξ0 = L. We then choose λ = α ⊗ β, where α = (1, 0, . . . , 0) ∈ Rn+1 and β = (1, 0, . . . , 0) ∈ Rn . We will see that LK (ξ0 + α ⊗ Bε , λ) is bounded for some ε > 0. Let η ∈ LK (ξ0 +α ⊗Bε , λ) then we can write η = ξ0 +α ⊗γε +tλ for some γε ∈ Bε and t ∈ R. By definition of LK (ξ0 + α ⊗ Bε , λ) we have adjn η ∈  S. Since S is bounded and   |adjn η| = λ1 + γε1 + t λ2 · · · λn , it follows, using the fact that rank ξ0
n − 1, that |t| is bounded by a constant depending on S, ξ0 and ε. Consequently |η|  |ξ0 | + |α ⊗ γε | + |t||λ| is bounded for any fixed positive ε and we get the result. 

As already alluded in Section 5.3, we obtain now a result of nonexistence although the integrand of the relaxed problem is not strictly quasiconvex. We will consider the case where N = 3, n = 2 and f : R3×2 → R is given by f (ξ ) = g(adj2 ξ ), where g : R3 → R is defined by

2 g(ν) = ν12 − 4 + ν22 + ν32 .

We therefore get Qf (ξ ) = Cg(adj2 ξ ) and

where

&2 % Cg(ν) = ν12 − 4 + + ν22 + ν32 ,

[x]+ =




x 0

if x
0, if x < 0.

We will choose the boundary datum as follows   1 uξ0 (x) = α1 x1 + α2 x2   u2ξ0 (x) = 0 uξ0 (x) =   u3ξ0 (x) = 0

and hence

Duξ0 (x) = ξ0 =



α1 0 0

α2 0 0



,

  0 adj2 Duξ0 (x) = adj2 ξ0 = 0 . 0

The problem is then




  1,∞ 3 inf I (u) = f Du(x) dx: u ∈ uξ0 + W0 Ω; R . Ω

(P)

120

B. Dacorogna

Note also that Qf (ξ0 ) = 0 < f (ξ0 ) = 16. In terms of the preceding notations we have   S = y ∈ R3 : Cg(y) < g(y) = y = (y1 , y2 , y3 ) ∈ R3 : |y1 | < 2 ,   K = ξ ∈ R3×2 : Qf (ξ ) < f (ξ ) = ξ ∈ R3×2 : adj2 ξ ∈ S

and we observe that Cg is not affine on S, which in turn implies that Qf is not quasiaffine on K. The following result shows that the hypothesis of strict quasiconvexity of Qf is not necessary for nonexistence. P ROPOSITION 78. (P) has a solution if and only if uξ0 ≡ 0. Moreover, Qf is not strictly quasiconvex at any ξ0 ∈ R3×2 of the form ξ0 =



α1 0 0

α2 0 0



.

P ROOF. Step 1. We first show that if (P) has a solution then uξ0 ≡ 0. If u ∈ uξ0 + W01,∞ (Ω; R3 ) is a solution of (P) we necessarily have, denoting by ν(ξ ) = adj2 ξ ,

since

  ν1 (Du) = 2,

ν2 (Du) = ν3 (Du) = 0

Qf (Duξ0 ) = Cg(adj2 Duξ0 ) = Cg(0) = 0. The three equations read as    2 3 u u − u2x2 u3x1  = 2,   x1 x2 u1x1 u3x2 − u1x2 u3x1 = 0,   1 2 ux1 ux2 − u1x2 u2x1 = 0.

(31)

Multiplying the second equation of (31) first by u2x1 , then by u2x2 , using the third equation of (31), we get

 0 = u2x1 u1x1 u3x2 − u2x1 u1x2 u3x1 = u2x1 u1x1 u3x2 − u1x1 u2x2 u3x1 = u1x1 u2x1 u3x2 − u2x2 u3x1 , 

0 = u2x2 u1x1 u3x2 − u2x2 u1x2 u3x1 = u2x1 u1x2 u3x2 − u2x2 u1x2 u3x1 = u1x2 u2x1 u3x2 − u2x2 u3x1 .

Combining these last equations with the first one of (31), we find u1x1 = u1x2 = 0

a.e.

Nonconvex problems of the calculus of variations and differential inclusions

121

We therefore find that any solution of (P) should have Du1 = 0 a.e. and hence u1 ≡ constant on each connected component of Ω. Since u1 agrees with u1ξ0 on the boundary of Ω, we deduce that u1ξ0 ≡ 0 and thus uξ0 ≡ 0, as claimed. Step 2. We next show that if uξ0 ≡ 0, then (P) has a solution. It suffices to choose u1 ≡ 0 and to solve 
 2 3 ux1 ux2 − u2x2 u3x1  = 2 a.e. in Ω, u2 = u3 = 0

on ∂Ω.

This is possible by virtue of, for example, Corollary 7.30 in [31]. Step 3. We finally prove that Qf is not strictly quasiconvex at any ξ0 ∈ R3×2 of the form given in the statement of the proposition. Indeed, let 0 < R1 < R2 < R and denote by BR the ball centered at 0 and of radius R. Choose λ, µ ∈ C ∞ (BR ) such that (1) λ = 0 on ∂BR and λ ≡ 1 on BR2 , R2 , µ ≡ 1 on BR1 and (2) µ ≡ 0 on BR B   2 µ + µ(x1 µx + x2 µx ) < 2 for every x ∈ BR . 1 2

R1 ) is easily ensured by choosing This last condition (which is a restriction only in BR2 B appropriately R1 , R2 and R. We then choose u(x) = uξ0 (x) + ϕ(x), where ϕ 1 (x) = −λ(x)u1ξ0 (x),

ϕ 2 (x) = µ(x)x1

and ϕ 3 (x) = µ(x)x2 .

R2 , while on BR2 we We therefore have that ϕ ∈ W01,∞ (BR ; R3 ), adj2 Du ≡ 0 on BR B have 

adj2 Du = µ2 + µ(x1 µx1 + x2 µx2 ), 0, 0 .

We have thus obtained that Cg(adj2 Du) ≡ 0 and hence Qf (ξ0 + Dϕ) ≡ Qf (ξ0 ) = 0. This implies that (QP) has infinitely many solutions. However since ϕ does not vanish identically, we deduce that Qf is not strictly quasiconvex at any ξ0 of the given form.  7.6. The problem of potential wells The general problem of potential wells has been intensively studied by many authors in conjunction with crystallographic models involving fine microstructures. The reference paper on the subject is Ball and James [8]. It has then been studied by many authors including Bhattacharya, Firoozye, James and Kohn, Dacorogna and Marcellini, De Simone

122

B. Dacorogna

and Dolzmann, Dolzmann and Müller, Ericksen, Firoozye and Kohn, Fonseca and Tartar, Kinderlehrer and Pedregal, Kohn, Luskin, Müller and Sverak, Pipkin, Sverak, and we refer to [31] for exact bibliographic references. In mathematical terms the problem of potential wells can be described as follows. Find a minimizer of the problem






 f Du(x) dx: u ∈ uξ0 + W01,∞ Ω; Rn , inf

(P)

Ω

where Ω ⊂ Rn is a bounded open set, uξ0 is an affine map with Duξ0 = ξ0 and f : Rn×n → R+ is such that f (ξ ) = 0

⇐⇒

ξ ∈E=

m *

SO(n)Ai .

i=1

The m wells are SO(n)Ai , 1  i  m (and SO(n) denotes the set of matrices U such that U ⊤ U = U U ⊤ = I and det U = 1). The interesting case is when ξ0 ∈ int Rco E, and we have then that Qf (ξ0 ) = 0. Therefore by the relaxation theorem we have inf(P) = inf(QP) = 0. The existence of minimizers, since Qf is affine on Rco E (indeed Qf ≡ 0), for (P) is then reduced to finding a function u ∈ uξ0 + W01,∞ (Ω; Rn ) so that Du(x) ∈ E =

m *

SO(n)Ai .

i=1

The problem is relatively well understood only in the cases of two wells, i.e. m = 2, and in dimension n = 2. It is this case that we briefly discuss now. We therefore have now A, B ∈ R2×2 with 0 < det A < det B and we want to find u ∈ uξ0 + W01,∞ (Ω; R2 ), where Ω ⊂ R2 is a bounded open set, satisfying Du(x) ∈ SO(2)A ∪ SO(2)B

a.e. in Ω.

Nonconvex problems of the calculus of variations and differential inclusions

123

The first important result is to identify the set where the gradient of the boundary datum, ξ0 , should lie. This was resolved by Sverak [73] who showed that
det B − det ξ , Rco E = ξ ∈ R2×2 : there exist 0  α  det B − det A det ξ − det A R, S ∈ SO(2), det B − det A

so that ξ = αRA + βSB 0β 

while the interior is given by the same formulas with strict inequalities in the right-hand side. We therefore have the following theorem. T HEOREM 79. Let Ω ⊂ R2 be a bounded open set ξ0 ∈ int Rco E. Then there exists u ∈ uξ0 + W01,∞ (Ω; R2 ) such that Du(x) ∈ E = SO(2)A ∪ SO(2)B

a.e. in Ω

and therefore (P) has a solution. This result was proved by Müller and Sverak [61] using the so-called method of convex integration of Gromov [43] and by Dacorogna and Marcellini in [28] and [31] following the approach presented in Section 4.4, and we refer to [31] for details. The case where det A = det B > 0 can also be handled (cf. Müller and Sverak [62], see also Dacorogna and Tanteri [37]), using the representation formula of Sverak [73], namely  Rco E = ξ ∈ R2×2 : there exist R, S ∈ SO(2), 0  α, β  α + β  1, det ξ = det A = det B so that ξ = αRA + βSB . References [1] E. Acerbi and N. Fusco, Semicontinuity problems in the calculus of variations, Arch. Ration. Mech. Anal. 86 (1984), 125–145. [2] J.J. Alibert and B. Dacorogna, An example of a quasiconvex function that is not polyconvex in two dimensions, Arch. Ration. Mech. Anal. 117 (1992), 155–166. [3] G. Allaire and G. Francfort, Existence of minimizers for nonquasiconvex functionals arising in optimal design, Ann. Inst. H. Poincaré Anal. Non Linéaire 15 (1998), 301–339. [4] G. Aubert and R. Tahraoui, Théorèmes d’existence pour des problèmes du calcul des variations, J. Differential Equations 33 (1979), 1–15. [5] G. Aubert and R. Tahraoui, Sur la minimisation d’une fonctionelle nonconvexe, non différentiable en dimension 1, Boll. Unione Mat. Ital. 17 (1980), 244–258.

124

B. Dacorogna

[6] G. Aubert and R. Tahraoui, Théorèmes d’existence en optimisation non convexe, Appl. Anal. 18 (1984), 75–100. [7] J.M. Ball, Convexity conditions and existence theorems in nonlinear elasticity, Arch. Ration. Mech. Anal. 63 (1977), 337–403. [8] J.M. Ball and R.D. James, Fine phase mixtures as minimizers of energy, Arch. Ration. Mech. Anal. 100 (1987), 15–52. [9] S. Bandyopadhyay, A.C. Barroso, B. Dacorogna and J. Matias, Differential inclusions for differential forms and applications to the calculus of variations, to appear. [10] P. Bauman and D. Phillips, A nonconvex variational problem related to change of phase, J. Appl. Math. Optim. 21 (1990), 113–138. [11] A. Bressan and F. Flores, On total differential inclusions, Rend. Sem. Mat. Univ. Padova 92 (1994), 9–16. [12] G. Buttazzo, B. Dacorogna and W. Gangbo, On the envelopes of functions depending on singular values of matrices, Boll. Unione Mat. Ital. Sez. B 8 (1994), 17–35. [13] G. Buttazzo, V. Ferone and B. Kawohl, Minimum problems over sets of concave functions and related questions, Math. Nachr. 173 (1995), 71–89. [14] P. Celada and S. Perrotta, Minimizing nonconvex multiple integrals: A density result, Proc. Roy. Soc. Edinburgh Sect. A 130 (2000), 721–741. [15] P. Celada and S. Perrotta, On the minimum problem for nonconvex multiple integrals of product type, Calc. Var. Partial Differential Equations 12 (2001), 371–398. [16] A. Cellina, On minima of a functional of the gradient: Necessary conditions, Nonlinear Anal. 20 (1993), 337–341. [17] A. Cellina, On minima of a functional of the gradient, sufficient conditions, Nonlinear Anal. 20 (1993), 343–347. [18] A. Cellina and G. Colombo, On a classical problem of the calculus of variations without convexity conditions, Ann. Inst. H. Poincaré Anal. Non Linéaire 7 (1990), 97–106. [19] A. Cellina and S. Zagatti, An existence result in a problem of the vectorial case of the calculus of variations, SIAM J. Control Optim. 33 (1995), 960–970. [20] L. Cesari, An existence theorem without convexity conditions, SIAM J. Control Optim. 12 (1974), 319–331. [21] L. Cesari, Optimization – Theory and Applications, Springer-Verlag, Berlin (1983). [22] A. Cutrì, Some remarks on Hamilton–Jacobi equations and nonconvex minimization problems, Rend. Mat. Appl. 13 (1993), 733–749. [23] B. Dacorogna, A relaxation theorem and its applications to the equilibrium of gases, Arch. Ration. Mech. Anal. 77 (1981), 359–386. [24] B. Dacorogna, Minimal hypersurfaces in parametric form with non convex integrands, Indiana Univ. Math. J. 31 (1982), 531–552. [25] B. Dacorogna, Quasiconvexity and relaxation of non convex variational problems, J. Funct. Anal. 46 (1982), 102–118. [26] B. Dacorogna, Direct Methods in the Calculus of Variations, Appl. Math. Sci., Vol. 78, Springer-Verlag, Berlin (1989). [27] B. Dacorogna and P. Marcellini, Existence of minimizers for non quasiconvex integrals, Arch. Ration. Mech. Anal. 131 (1995), 359–399. [28] B. Dacorogna and P. Marcellini, Théorèmes d’existence dans le cas scalaire et vectoriel pour les équations de Hamilton–Jacobi, C. R. Acad. Sci. Paris 322 (1996), 237–240. [29] B. Dacorogna and P. Marcellini, General existence theorems for Hamilton–Jacobi equations in the scalar and vectorial case, Acta Math. 178 (1997), 1–37. [30] B. Dacorogna and P. Marcellini, On the solvability of implicit nonlinear systems in the vectorial case, Nonlinear Partial Differential Equations, G.-Q. Chen and E. DiBenedetto, eds, Contemp. Math., Vol. 238, Amer. Math. Soc., Providence, RI (1999), 89–113. [31] B. Dacorogna and P. Marcellini, Implicit Partial Differential Equations, Birkhäuser, Boston (1999). [32] B. Dacorogna and P. Marcellini, Attainment of minima and implicit partial differential equations, Ricerche Mat. 48 (1999), 311–346. [33] B. Dacorogna and G. Pisante, A general existence theorem for differential inclusions in the vector valued case, Portugal. Math., to appear.

Nonconvex problems of the calculus of variations and differential inclusions

125

[34] B. Dacorogna, G. Pisante and A.M. Ribeiro, On non quasiconvex problems of the calculus of variations, Discrete Contin. Dyn. Syst. Ser. A, to appear. [35] B. Dacorogna and A.M. Ribeiro, Existence of solutions for some implicit pdes and applications to variational integrals involving quasiaffine functions, Proc. Roy. Soc. Edinburgh Sect. A 134 (2004), 1–15. [36] B. Dacorogna and C. Tanteri, On the different convex hulls of sets involving singular values, Proc. Roy. Soc. Edinburgh Sect. A 128 (1998), 1261–1280. [37] B. Dacorogna and C. Tanteri, Implicit partial differential equations and the constraints of nonlinear elasticity, J. Math. Pures Appl. 81 (2002), 311–341. [38] F.S. De Blasi and G. Pianigiani, On the Dirichlet problem for Hamilton–Jacobi equations. A Baire category approach, NoDEA Nonlinear Differential Equations Appl. 6 (1999), 13–34. [39] I. Ekeland, Nonconvex minimization problems, Bull. Amer. Math. Soc. 1 (1979), 443–475. [40] G. Friesecke, A necessary and sufficient condition for non attainment and formation of microstructure almost everywhere in scalar variational problems, Proc. Roy. Soc. Edinburgh Sect. A 124 (1994), 437–471. [41] N. Fusco, P. Marcellini and A. Ornelas, Existence of minimizers for some nonconvex one-dimensional integrals, Portugal. Math. 55 (1998), 167–185. [42] D. Giachetti and R. Schianchi, Minima of some nonconvex noncoercive problems, Ann. Mat. Pura Appl. 165 (1993), 109–120. [43] M. Gromov, Partial Differential Relations, Springer-Verlag, Berlin (1986). [44] R.A. Horn and C.A. Johnson, Matrix Analysis, Cambridge University Press, Cambridge (1985). [45] R.A. Horn and C.A. Johnson, Topics in Matrix Analysis, Cambridge University Press, Cambridge (1991). [46] B. Kirchheim, Deformations with finitely many gradients and stability of quasiconvex hulls, C. R. Acad. Sci. Paris 332 (2001), 289–294. [47] R. Klötzler, On the existence of optimal processes, Banach Center of Warsaw Publications (1976), 125–130. [48] R.V. Kohn and G. Strang, Optimal design and relaxation of variational problems I, II, III, Comm. Pure Appl. Math. 39 (1986), 113–137, 139–182, 353–377. [49] H. Le Dret and A. Raoult, Enveloppe quasi-convexe de la densité d’énergie de Saint Venant–Kirchhoff, C. R. Acad. Sci. Paris 318 (1994), 93–98. [50] P. Marcellini, Alcune osservazioni sull’esistenza del minimo di integrali del calcolo delle variazioni senza ipotesi di convessità, Rend. Mat. 13 (1980), 271–281. [51] P. Marcellini, A relation between existence of minima for nonconvex integrals and uniqueness for not strictly convex integrals of the calculus of variations, Mathematical Theories of Optimization, J.P. Cecconi and T. Zolezzi, eds, Lecture Notes in Math., Vol. 979, Springer-Verlag, Berlin (1983), 216–231. [52] P. Marcellini, Non convex integrals of the calculus of variations, Methods of Nonconvex Analysis, A. Cellina, ed., Lecture Notes in Math., Vol. 1446, Springer-Verlag, Berlin (1990), 16–57. [53] A.W. Marshall and I. Olkin, Inequalities: Theory of Majorisation and Its Applications, Academic Press, New York (1979). [54] E. Mascolo, Some remarks on nonconvex problems, Material Instabilities in Continuum Mechanics, J.M. Ball, ed., Oxford University Press, Oxford (1988), 269–286. [55] E. Mascolo and R. Schianchi, Existence theorems for nonconvex problems, J. Math. Pures Appl. 62 (1983), 349–359. [56] E. Mascolo and R. Schianchi, Nonconvex problems in the calculus of variations, Nonlinear Anal. 9 (1985), 371–379. [57] E. Mascolo and R. Schianchi, Existence theorems in the calculus of variations, J. Differential Equations 67 (1987), 185–198. [58] M.D.P. Monteiro Marques and A. Ornelas, Genericity and existence of a minimum for scalar integral functionals, J. Optim. Theory Appl. 86 (1995), 421–431. [59] C.B. Morrey, Quasiconvexity and the lower semicontinuity of multiple integrals, Pacific J. Math. 2 (1952), 25–53. [60] C.B. Morrey, Multiple Integrals in the Calculus of Variations, Springer-Verlag, Berlin (1966). [61] S. Müller and V. Sverak, Attainment results for the two-well problem by convex integration, Geometric Analysis and the Calculus of Variations, J. Jost, ed., International Press, Cambridge, MA (1996), 239–251. [62] S. Müller and V. Sverak, Convex integration with constraints and applications to phase transitions and partial differential equations, J. Eur. Math. Soc. 1 (1999), 393–422.

126

B. Dacorogna

[63] S. Müller and M. Sychev, Optimal existence theorems for nonhomogeneous differential inclusions, J. Funct. Anal. 181 (2001), 447–475. [64] A. Ornelas, Existence of scalar minimizers for nonconvex simple integrals of sum type, J. Math. Anal. Appl. 221 (1998), 559–573. [65] J.C. Oxtoby, Measure and Category, Springer-Verlag, Berlin (1971). [66] G. Pisante, Implicit partial differential equations: Different approaching methods, Ph.D. Thesis, Naples (2004). [67] J.P. Raymond, Champs Hamiltoniens, relaxation et existence de solutions en calcul des variations, J. Differential Equations 70 (1987), 226–274. [68] J.P. Raymond, Conditions nécessaires et suffisantes de solutions en calcul des variations, Ann. Inst. H. Poincaré Anal. Non Linéaire 4 (1987), 169–202. [69] J.P. Raymond, Théorème d’existence pour des problèmes variationnels non convexes, Proc. Roy. Soc. Edinburgh Sect. A 107 (1987), 43–64. [70] J.P. Raymond, Existence of minimizers for vector problems without quasiconvexity conditions, Nonlinear Anal. 18 (1992), 815–828. [71] R.T. Rockafellar, Convex Analysis, Princeton University Press, Princeton, NJ (1970). [72] V. Sverak, Rank one convexity does not imply quasiconvexity, Proc. Roy. Soc. Edinburgh Sect. A 120 (1992), 185–189. [73] V. Sverak, On the problem of two wells, Microstructure and Phase Transitions, J. Ericksen et al., eds, IMA Vol. Appl. Math., Vol. 54, Springer-Verlag, Berlin (1993), 183–189. [74] M. Sychev, Comparing two methods of resolving homogeneous differential inclusions, Calc. Var. Partial Differential Equations 13 (2001), 213–229. [75] M.A. Sychev, Characterization of homogeneous scalar variational problems solvable for all boundary data, Proc. Roy. Soc. Edinburgh Sect. A 130 (2000), 611–631. [76] R. Tahraoui, Théorèmes d’existence en calcul des variations et applications à l’élasticité non linéaire, C. R. Acad. Sci. Paris 302 (1986), 495–498; Proc. Roy. Soc. Edinburgh Sect. A 109 (1988), 51–78. [77] R. Tahraoui, Sur une classe de fonctionnelles non convexes et applications, SIAM J. Math. Anal. 21 (1990), 37–52. [78] G. Treu, An existence result for a class of nonconvex problems of the calculus of variations, J. Convex Analysis 5 (1998), 31–44. [79] K. Yosida, Functional Analysis, 6th Edition, Springer-Verlag, Berlin (1980). [80] S. Zagatti, Minimization of functionals of the gradient by Baire’s theorem, SIAM J. Control Optim. 38 (2000), 384–399.

CHAPTER 3

Bifurcation and Related Topics in Elliptic Problems Yihong Du School of Mathematics, Statistics and Computer Science, University of New England, Armidale, NSW 2351, Australia E-mail: [email protected]

Contents 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2. Bifurcation from infinity by spatial degeneracy . . . . . . . . . . . . 2.1. Bifurcation from infinity and boundary blow-up problems . . . 2.2. Perturbation and patterned solutions . . . . . . . . . . . . . . . 2.3. Comments and related results . . . . . . . . . . . . . . . . . . 3. Bifurcation and monotonicity: A heterogeneous competition system 3.1. Global bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . 3.2. Stability analysis . . . . . . . . . . . . . . . . . . . . . . . . . 3.3. Stable patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4. Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4. Bifurcation and exact multiplicity: The perturbed Gelfand equation 4.1. The limiting equations . . . . . . . . . . . . . . . . . . . . . . 4.2. The perturbed Gelfand equation in dimensions 1 and 2 . . . . . 4.3. The perturbed Gelfand equation in higher dimensions . . . . . 4.4. Further remarks and related results . . . . . . . . . . . . . . . . 5. Nodal properties and global bifurcation . . . . . . . . . . . . . . . . 5.1. Global bifurcation of the competition system . . . . . . . . . . 5.2. Uniqueness results for the predator–prey system . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

HANDBOOK OF DIFFERENTIAL EQUATIONS Stationary Partial Differential Equations, volume 2 Edited by M. Chipot and P. Quittner © 2005 Elsevier B.V. All rights reserved 127

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

129 131 132 137 144 147 147 154 163 168 169 171 182 188 192 193 193 200 205

Bifurcation and related topics

129

1. Introduction Bifurcation theory provides a bridge between the linear world and the more complicated nonlinear world, and thus plays an important role in the study of various nonlinear problems. Nonlinear elliptic boundary value problems enjoy many nice properties that allow the use of a variety of powerful tools in nonlinear functional analysis. In the past three decades, bifurcation theory has been successfully combined with these tools to yield rather deep results for elliptic problems. Traditionally bifurcation analysis was based on local linearization techniques, but more and more global analysis is involved in modern bifurcation theory. A highlight is the global bifurcation theory of Krasnoselskii and Rabinowitz (see [Kr,Ra]), which is resulted from the use of topological degree theory, general set point theory and a linearization consideration. The final result, generally known as Rabinowitz’s global bifurcation theorem (to be recalled later), has played a fundamental role in proving a great number of existence results for elliptic problems. Making use of the maximum principle, one can study elliptic problems in the framework of ordered Banach spaces. The extra order structure greatly strengthens these abstract tools. An excellent presentation of these techniques up to the late 1970s can be found in Amann’s by now classical review article [Am]. In this chapter we intend to present some further results for elliptic boundary value problems, where bifurcation theory plays an important role in the proofs; we will focus on recent developments, well after [Am]. In Section 2 we discuss a new phenomenon, namely bifurcation from infinity caused by spatial “degeneracy” in the nonlinearity and determined by “boundary blow-up solutions”. In Section 3 we combine bifurcation argument and order structure to study a system of elliptic equations, and demonstrate that, apart from multiplicity results, these techniques can be used to discuss the stability and the profiles of the solutions. In Section 4 we present some recent fine techniques in determining the exact number of positive solutions of elliptic equations over a ball; in particular, we give the proof of a long standing conjecture on the perturbed Gelfand equation. In Section 5 we discuss the usefulness of nodal properties of solutions in global bifurcation theory. Most of the problems discussed here have an open ending; related problems and open questions can be found in the remarks at the end of the sections or subsections. The choice of the topics in this chapter is subjective, and the bibliography is by no means complete. For clarity and simplicity, we present most of our results for problems with a specific nonlinearity, with the hope that the interested reader can easily extend them to more general situations. We have not tried to make this chapter entirely self-contained. Most of the results presented here are proved in full, but some of them are only stated, with the proofs referred to the relevant references. The proofs not given here are either very technical or less relevant to our main theme here. We assume that the reader is familiar with the standard theory for second-order linear elliptic equations (see [GT] and [PW]) and standard nonlinear functional analysis (see [De]). Some classical bifurcation theorems. We recall several classical bifurcation theorems which form the corner stones for our analysis in this chapter; indeed, they are fundamental in the development of the modern bifurcation theory in general.

130

Y. Du

We first describe Rabinowitz’s global bifurcation theorem. Let E be a Banach space and R1 denote the set of real numbers. We consider the nonlinear eigenvalue problem u = λLu + H (λ, u),

(λ, u) ∈ R1 × E,

(1.1)

where L : E → E is a compact linear map, H : R1 × E → E is compact and continuous, and is o(u) for u near 0 uniformly on bounded λ intervals. We assume that H (λ, 0) ≡ 0, and therefore we have the curve of trivial solutions {(λ, 0): λ ∈ R1 }. We are interested in the existence of nontrivial solutions (λ, u) ∈ R1 × E and will denote the closure of the set of nontrivial solutions of (1.1) by S. Let r(L) denote the set of µ ∈ R1 such that there exists v ∈ E \ {0} with v = µLv. It is well known that the possible bifurcations points for (1.1) with respect to the curve of trivial solutions lie in the set {(µ, 0): µ ∈ r(L)}; moreover, if µ ∈ r(L) is of odd (algebraic) multiplicity, then (µ, 0) is a bifurcation point, see [Kr]. T HEOREM 1.1 (Rabinowitz’s global bifurcation theorem [Ra]). Under the above assumptions, if µ ∈ r(L) is of odd (algebraic) multiplicity, then S possesses a maximal subcontinuum Sµ such that (µ, 0) ∈ Sµ and Sµ either (i) meets infinity in R1 × E, i.e., Sµ is unbounded, or (ii) meets (µ, ˆ 0), where µˆ ∈ r(L) \ {µ}. Next we recall two local bifurcation theorems due to Crandall and Rabinowitz [CR1,CR2]. Let E1 , E2 be Banach spaces and J = (a, b) an open interval in R1 . Let N (L) and R(L) denote the null space and range of a linear map L between Banach spaces. T HEOREM 1.2 (Bifurcation from a simple eigenvalue [CR1]). Let E1 , E2 and J be as above. Suppose that U is a neighborhood of 0 in E1 , λ0 ∈ J and F : J × U → E2 has the following properties: (a) F (λ, 0) ≡ 0 for λ ∈ J , (b) the partial derivatives Fλ , Fu and Fλu exist and are continuous, (c) dim N (Fu (λ0 , 0)) = codim R(Fu (λ0 , 0)) = 1, (d) Fλu (λ0 , 0)u0 ∈ / R(Fu (λ0 , 0)), where u0 ∈ E1 spans N (Fu (λ0 , 0)). Let Z be any complement of span{u0 } in E1 . Then there exists an open interval J0 containing 0 and continuously differentiable functions λ : J0 → R1 and ψ : J0 → Z such that λ(0) = λ0 , ψ(0) = 0, and if u(s) = su0 + sψ(s), then F (λ(s), u(s)) = 0. Moreover, the solution set of F (λ, u) = 0 near (λ0 , 0) consists precisely of the curves {(λ(s), u(s)): s ∈ J0 } and {(λ, 0): λ ∈ J }. If the equation F (λ, u) = 0 in Theorem 1.2 can be written in the form of (1.1), then conditions (c) and (d) of Theorem 1.2 become dim N (λ0 L − I ) = codim R(λ0 L − I ) = 1 and u0 ∈ / R(λ0 L − I )

if N (λ0 L − I ) = span{u0 }.

Bifurcation and related topics

131

If λ0 ∈ r(L) has the above properties, then one says that 1/λ0 is a simple eigenvalue of L. Therefore Theorem 1.2 is usually known as the theorem of bifurcation from a simple eigenvalue; it provides a much better description of the local bifurcation branch. Both Theorems 1.1 and 1.2 describe the situation that a nontrivial solution branch bifurcates from a trivial solution curve. The following theorem describes the situation that a solution curve “changes direction” in the (λ, u) space. As will be discussed later, this theorem plays a vital role in obtaining exact multiplicity results. T HEOREM 1.3 (Turning point theorem [CR2]). Let E1 , E2 and J be as in Theorem 1.2. Suppose that V is a neighborhood of v0 in E1 , λ0 ∈ J and F : J × V → E2 is continuously differentiable and has the following properties: (a) F (λ0 , v0 ) = 0, (b) dim N (Fu (λ0 , v0 )) = codim R(Fu (λ0 , v0 )) = 1, (c) Fλ (λ0 , v0 ) ∈ / R(Fu (λ0 , v0 )). Let Z be any complement of span{u0 } in E1 , where u0 ∈ E1 spans N (Fu (λ0 , v0 )). Then there exists an open interval J0 containing 0 and continuously differentiable functions λ : J0 → R1 and τ : J0 → Z such that λ(0) = λ0 , λ′ (0) = 0, τ (0) = τ ′ (0) = 0, and if u(s) = v0 + su0 + τ (s), then F (λ(s), u(s)) = 0. Moreover, the solution set of F (λ, u) = 0 near (λ0 , v0 ) consists precisely of the curve {(λ(s), u(s)): s ∈ J0 }. Furthermore, if F is k-times continuously differentiable (analytic), so are λ(s) and τ (s). Note that if we can somehow determine the sign of λ′′ (0) in Theorem 1.3, then we would know in which direction the solution curve is bent near (λ0 , v0 ). For example, if λ′′ (0) < 0, then the solution curve is bent to the left, i.e., for values of λ less than λ0 . In the case λ′′ (0) = 0, it is possible that the solution curve does not change direction at (λ0 , v0 ) but behaves like the curve x = y 3 at (0, 0) in the xy-plane. 2. Bifurcation from infinity by spatial degeneracy In this section we use the problem −
u = λu − b(x)|u|p−1 u,

u|∂Ω = 0,

(2.1)

to demonstrate how bifurcation from infinity can be caused by b(x) vanishing in a subset of the underlying domain Ω. We call this behavior of b(x) a degeneracy in (2.1). Here Ω is a bounded smooth domain in RN , p > 1 and b(x) is a continuous nonnegative function  Positive solutions of problem (2.1) can be regarded as steady-states of a biological over Ω. species over the spatial region Ω, whose growth is governed by a degenerate logistic law. When b(x) is replaced by b(x) + ε with a small positive constant ε, (2.1) describes the steady-states of a species governed by a classical logistic law. We will make use of (2.1) to study the perturbed problem and reveal that, for small ε, the profile (or pattern) of the positive solutions for the perturbed problem can be determined rather completely. This is of interest in population biology; we will also use these results in Section 3.

132

Y. Du

2.1. Bifurcation from infinity and boundary blow-up problems We consider positive solutions of (2.1); as will become clear soon, for this case, the theory is rather complete. We note that by the strong maximum principle, a nontrivial nonnegative 0 := b−1 (0), where solution of (2.1) must be strictly positive inside Ω. We assume that Ω Ω0 is an open connected set whose closure is contained in Ω, and both ∂Ω and ∂Ω0 are Ω0 smooth (say, C 2 ). We denote by λΩ 1 and λ1 the first eigenvalues of −
under Dirichlet boundary conditions over Ω and Ω0 , respectively. In general, we use λD 1 (φ) to denote the first eigenvalue of −
+ φ over D under Dirichlet boundary conditions. If (2.1) has a positive solution, then we write (2.1) in the form

 −
u + b(x)up−1 u = λu p−1 ) = λ. Using the well-known monotonicity properties of and obtain that λΩ 1 (b(x)u D λ1 (φ), we deduce

  Ω Ω Ω p−1 λΩ < λ1 0 b(x)up−1 = λ1 0 1 < λ1 b(x)u

since b(x) = 0 on Ω0 . Therefore we have the necessary condition, Ω

0 λΩ 1 < λ < λ1 ,

(2.2)

for (2.1) to possess a positive solution. We claim that (2.2) is also a sufficient condition for the existence of a positive solution of (2.1). This can be proved by combining a global bifurcation argument with an a priori bound result. Indeed, by a standard application of Rabinowitz’s global bifurcation result, we know that there is a global branch of positive solutions Γ := {(λ, u)} bifurcating from the trivial solution curve {(λ, 0)} at (λΩ 1 , 0). By the maximum principle and the fact that λΩ 1 is a simple eigenvalue, we conclude that the second alternative in Theorem 1.1 cannot occur and hence Γ has to be unbounded in the  space R1 × C 1 (Ω). L EMMA 2.1. For any small δ > 0, there exists C = Cδ > 0 such that any positive solution Ω of (2.1) with λ  λ1 0 − δ satisfies uC 1 (Ω)   C. P ROOF. By the standard Lp -theory for elliptic equations and the Sobolev embedding theorems, we only need to show the bound in the L∞ (Ω) norm. Indeed, if we have the bound for u in L∞ (Ω), then the right-hand side of (2.1) has a bound in L∞ (Ω). By the Lp -theory, this implies that u has a bound in W 2,q (Ω) for any q > 1. Then the Sobolev embedding  theorem shows that u has a bound in C 1 (Ω). From the continuous dependence of λD 1 on D,σ we can find a small neighborhood of Ω0 , Ω0 σ 0 say Ω0σ := {x ∈ Ω: d(x, Ω0 ) < σ }, so that λΩ 1 > λ1 − δ. Let φ be a positive eigenσ Ω σ  function corresponding to λ1 0 . Then extend φ |Ω0σ/2 to a smooth positive function over Ω, which we denote by φ. Then it is easily checked that there exists M0 > 0 large such that Ω0 for any λ ∈ (λΩ 1 , λ1 − δ] and all M
M0 , Mφ is an upper solution of (2.1). By the wellknown Serrin’s sweeping principle (see, e.g., Theorem 2.7.1 in [Sa]), we can conclude that

Bifurcation and related topics

133

any positive solution of (2.1) satisfies u  M0 φ in Ω. This proves the a priori bound in the  L∞ (Ω) norm, as required. By Lemma 2.1 (and its proof), we see that the global branch Γ of positive solutions of (2.1) can become unbounded only through a sequence {(λn , un )} ⊂ Γ satisfying Ω λn → λ1 0 and un ∞ → ∞. In particular, we have proved that (2.1) has a positive solution if and only if (2.2) holds. Moreover, (λΩ 1 , 0) is a point of bifurcation from 0, and Ω0 (λ1 , ∞) is a point of bifurcation from infinity. Taking advantage of the special nonlinearity in (2.1), we can show that it has a unique positive solution when (2.2) holds. Indeed, for any fixed λ satisfying (2.2), we can use the family of upper solutions constructed in the proof of Lemma 2.1 and a standard iteration technique to conclude that (2.2) has a maximal positive solution u∗ . If u is any other positive solution to (2.1), then applying the strong maximum principle to the equation satisfied by u∗ − u we find that u∗ − u > 0 in Ω. Therefore we have



∗ p−1  p−1 > λΩ = λ. λ = λΩ 1 b(x) u 1 b(x)u

This contradiction proves the uniqueness. For each λ satisfying (2.2), if we denote the unique positive solution of (2.1) by uλ , then Ω0 the global bifurcation branch Γ can be expressed as Γ = {(λ, uλ ): λΩ 1 < λ < λ1 }. Near Ω the end point (λ1 , 0) of Γ , the local bifurcation theorem, Theorem 1.2, gives a detailed description of the behavior of Γ ; here one relies on a linear eigenvalue problem. To better Ω understand Γ near its other end point (λ1 0 , ∞), instead of a linear problem, we need the following nonlinear boundary blow-up problem −
u = λu − b(x)up

0 , in Ω \ Ω

u|∂Ω0 = ∞, u|∂Ω = 0.

(2.3)

Here by u|∂Ω0 = ∞, we mean u(x) → ∞

when d(x, ∂Ω0 ) → 0.

We have the following result (see [DH], Theorem 2.4). P ROPOSITION 2.2. For any λ ∈ R1 , problem (2.3) has at least one positive solution. Moreλ and a minimal positive solution U λ in the sense over, it has a maximal positive solution U λ . that any other positive solution u satisfies U λ  u  U 0 R EMARK 2.3. If there exist constants α
0 and β2 > β1 > 0 such that, for all x ∈ Ω \ Ω near ∂Ω0 , % &α % &α β1 d(x, ∂Ω0 )  b(x)  β2 d(x, ∂Ω0 ) ,

then it is proved in [Du6], Theorem 3.2, that (2.3) has a unique positive solution. Whether the positive solution of (2.3) is unique without any extra condition on b(x) is an open problem.

134

Y. Du Ω

The following result gives a rather complete description of uλ for λ close to λ1 0 . Ω

T HEOREM 2.4. Denote λ0 := λ1 0 . Then 0 as λ increases to λ0 , (i) uλ → ∞ uniformly on Ω \Ω 0 as λ increases to λ0 . (ii) uλ → U λ0 uniformly in any compact subset of Ω P ROOF. We first observe that uλ increases as λ increases. This follows from a simple upper and lower solution consideration, together with the uniqueness of uλ . Suppose Ω0 ′ λΩ 1 < λ < λ < λ1 . Then uλ′ is an upper solution of (2.1). If φ is a positive eigenfunction corresponding to λΩ 1 , then it is easy to see that εφ < uλ′ and is a lower solution to (2.1) for all small positive ε. Therefore (2.1) has a positive solution u satisfying εφ < u < uλ′ . We must have u = uλ as (2.1) has a unique positive solution. Therefore uλ < uλ′ . The monotonicity of uλ in λ implies that if we can prove (i) and (ii) along a sequence λn → λ0 , then the same conclusions hold for λ → λ0 . Let {δn } be a sequence of positive numbers decreasing to 0 such that  Ωn := x ∈ Ω: d(x, Ω0 ) < δn ⊂⊂ Ω

∀n
1.

n Ω Denote λn = λΩ 1 . Then λ1 < λn < λ0 and λn increases to λ0 as n → ∞. To simplify notation, we write un = uλn . The proof below is divided into four steps.

S TEP 1. un (x) → ∞ uniformly in any compact subset of Ω0 . Ω

Let φ0 be the positive eigenfunction corresponding to λ0 = λ1 0 normalized by φ0 ∞ = 1. Let K be an arbitrarily given compact subset of Ω0 . Define α0 := inf u1 (x),

β0 := min φ0 (x).

x∈Ω0

x∈K

Clearly α0 > 0,

β0 > 0,

un (x)
u1 (x)
α0

∀x ∈ Ω0 , ∀n
1.

(2.4)

Given any large number M > 0, we can find an open connected set K ∗ satisfying K ⊂ K ∗ ⊂⊂ Ω0 such that φ0 (x)
β0 2

∀x ∈ K.

(2.6)

Recall that b(x) = 0 on K ∗ . Hence un and (M/β0 )φn satisfy the same equation −
u = λn u. It now follows from (2.4) and (2.6) that (M/β0 )φn and un are, respectively, lower and upper solutions of the problem −
u = λn u

in K ∗ ,

u|∂K ∗ = α0 .

∗

As λn < λ0 < λK 1 , it follows from the maximum principle that, for all large n, un (x)


M M φn (x)
β0 2

∀x ∈ K ⊂ K ∗ .

Since M > 0 is arbitrary, this shows limn→∞ un (x) = ∞ uniformly in K. This proves Step 1. Since ∂Ω0 is C 2 , it satisfies a uniform interior ball condition: There exists R > 0 such 0 and Bx ∩ ∂Ω0 = {x}. that for any x ∈ ∂Ω0 , there is a ball Bx of radius R such that Bx ⊂ Ω S TEP 2. Let xn ∈ ∂Ω0 be such that un (xn ) = min un (x). x∈∂Ω0

If {un (xn )} is bounded, then we can find a constant σ > 0 and a sequence cn → ∞ such that un (x)
un (xn ) + cn ψ(x), 2

whenever

R  |x − yn |  R, 2

(2.7)

2

where ψ(x) = e−σ |x−yn | − e−σ R and yn is the center of the ball Bxn . A simple calculation gives

 2 2
ψ + λn ψ = 4σ 2 |x − yn |2 − 2N σ + λn e−σ |x−yn | − λn e−σ R .

We can choose a large σ > 0 such that −
ψ(x)  λn ψ(x)

∀x ∈ Bxn \ BR/2 (yn ),

where BR/2 (yn ) = {x ∈ RN : |x − yn | < R/2}.  Choose a compact set K ⊂⊂ Ω0 such that K ⊃ ∞ n=1 BR/2 (yn ). By Step 1 and the assumption that {un (xn )} is bounded, we can find a sequence cn → ∞ such that

2 2 un (x)
un (xn ) + cn e−σ R /4 − e−σ R

∀x ∈ BR/2 (yn ) ⊂ K.

136

Y. Du

On the other hand, since λn < λ0 , by the maximum principle, un (x)
un (xn ) ∀x ∈ Ω0 . In particular, un (x)
un (xn ) on ∂Bxn . Thus we see that un is an upper solution to the problem 

−
u = λn u

R/2 (yn ), in Bxn \ B

u|∂Bxn = un (xn ),

2 2 u|∂BR/2 (yn ) = un (xn ) + cn e−σ R /4 − e−σ R .

(2.8)

But clearly, un (xn ) + cn ψ(x) is a lower solution to (2.8). Hence, since λn < λ0 < R/2 (yn )) (Bxn \B

λ1

, by the maximum principle,

un (x)
un (xn ) + cn ψ(x),

whenever

R  |x − yn |  R, 2

as required. 0 . S TEP 3. limn→∞ un (x) = ∞ uniformly on Ω By the maximum principle, it suffices to show that un (xn ) = min un (x) → ∞. x∈∂Ω0

We argue indirectly. Suppose that this is not true. Then by passing to a subsequence, we may assume that {un (xn )} is bounded: un (xn )  C for all n. Clearly un is an upper solution to −
u = λn u − b∗ up

0 , in Ω \ Ω

u|∂Ω0 = un (xn ), u|∂Ω = 0,

(2.9)

where b∗ = b∞ . Since 0 is a lower solution, we see that (2.9) has a positive solution vn  un . Replacing un (xn ) in (2.9) by its upper bound C, we similarly obtain a positive solution V of (2.9) satisfying vn  V on Ω \ Ω0 . In particular, vn L∞ (Ω\Ω0 ) is bounded. Then the Lp estimates and the Sobolev embedding theorems imply that {vn } is bounded in  \ Ω0 ). In particular, |∇vn (xn )| is bounded. Since C 1 (Ω un (x)
vn (x) ∀x ∈ Ω \ Ω0

and un (xn ) = vn (xn ),

we have ∂un (xn ) ∂vn (xn )   C0 ∂νn ∂νn for some C0 > 0, where νn = (yn − xn )/|yn − xn |, and yn is as in Step 2. On the other hand, by Step 2, % ∂un (xn ) ∂ψ(xn ) 2&
cn = cn 2σ Re−σ R → ∞ ∂νn ∂νn

Bifurcation and related topics

137

as n → ∞. This contradiction finishes the proof of Step 3. λ0 in C 1 (K) as n → ∞. \Ω 0 , un → U S TEP 4. For any compact set K ⊂ Ω Here we need the following comparison result. L EMMA 2.5. Suppose that D is a bounded domain, α(x) and β(x) are continuous functions in D with α∞ < ∞, and β(x) is nonnegative and not identically zero. Let u1 , u2 ∈ C 1 (D) be positive in D and satisfy, in the weak sense, Lu1 + α(x)u1 − β(x)g(u1 )  0  Lu2 + α(x)u2 − β(x)g(u2 ),

x ∈ D,

and lim (u2 − u1 )  0,

x→∂D

 where Lu = ij [aij (x)uxi ]xj is a uniformly elliptic operator with smooth coefficients aij , and g(u) is continuous and such that g(u)/u is strictly increasing for u in the range infD {u1 , u2 } < u < supD {u1 , u2 }. Then u2  u1 in D. Lemma 2.5 was proved in [DM], Lemma 2.1, when u1 and u2 are C 2 ; the same proof works if they are only C 1 . Applying Lemma 2.5 we find that un  un+1  U λ0 . Hence limn→∞ un (x) = u∞ (x) exists and u∞ (x)  U λ0 . It follows that u∞ satisfies (2.3) with λ = λ0 . Here the fact that u∞ = ∞ on ∂Ω0 follows from un (x)  un+1 (x) and un (x) → ∞ uniformly on ∂Ω0 by Step 3. Since U λ0 is the minimal solution, we necessarily have u∞ = U λ0 . Using Sobolev embedding theorems and Lp estimates, we easily see that un → U λ0 in 1 \Ω 0 . This proves Step 4 and hence finishes C (K) as n → ∞, for any compact set K ⊂ Ω our proof of Theorem 2.4. 

2.2. Perturbation and patterned solutions If we replace b(x) by b(x) + ε in (2.1), we will see that its global bifurcation branch of positive solutions Γε differs considerably from Γ , no matter how small is the positive constant ε. In this subsection we examine closely the evolution of Γε as ε decreases to 0. We will see that some solutions on Γε develop a sharp pattern as ε → 0, others do not. So we consider the problem % & −
u = λu − b(x) + ε up ,

u|∂Ω = 0.

(2.10)

As is well known, a standard global bifurcation consideration can be applied to (2.10) to yield an unbounded global branch of positive solutions Γε := {(λ, u)}, bifurcating from the trivial solution curve at (λΩ 1 , 0). Moreover, (2.10) can have a positive Ω solution only if λ > λ1 ; this can be proved by the same trick used to prove (2.2). For any

138

Y. Du

given Λ > 0 and all λ  Λ, we can find M := MΛ > 0 large such that any constant C
M is an upper solution of (2.10). Analogous to Lemma 2.1, this gives an a priori bound for all positive solutions of (2.10) with λ < Λ. Therefore Γε can only become unbounded through λ → ∞. Furthermore, (2.10) has a unique positive solution uελ for any λ > λΩ 1 , which can be proved by the same argument used for (2.1). Therefore Γε =

  λ, uελ : λ > λΩ 1 .

The following result describes how Γε evolves as ε → 0. T HEOREM 2.6. Let uλ and uελ be the unique positive solutions to (2.1) and (2.10), respectively. Then the following hold. Ω0 ε  (i) If λΩ 1 < λ < λ1 , then uλ → uλ uniformly on Ω as ε → 0. Ω0 (ii) If λ
λ1 , then 0 as ε → 0, (a) uελ → ∞ uniformly on Ω \Ω 0 as ε → 0. (b) uελ → U λ uniformly on compact subsets of Ω P ROOF. Recall that by an upper and lower solution consideration and the uniqueness of uλ , we deduced that uλ (x) is increasing in λ. The same consideration can be used to show that uελ (x) is increasing in λ, decreasing in ε, and uελ (x) < uλ (x)

(2.11)

whenever both exist. (One can also apply Lemma 2.5 to prove these properties.) Ω0 ε Suppose now λΩ 1 < λ < λ1 . Then by (2.11), we know that {uλ : ε > 0} is bounded ∞ in L (Ω). By elliptic regularity and the Sobolev embedding theorem, we see that {uελ :  Since ε → uε (x) is decreasing, (2.11) implies that u0 (x) := ε > 0} is compact in C 1 (Ω). λ λ ε limε→0 uλ (x) exists for all x ∈ Ω. The above compactness conclusion then implies that  and furthermore, u0 is a positive solution of (2.1). Therefore we must uελ → u0λ in C 1 (Ω) λ 0 have uλ = uλ , due to uniqueness. This proves conclusion (i) of the theorem. Ω We next prove conclusion (ii). So we assume that λ
λ1 0 . Let mε = min uε (x) = uε (xε ), 0 x∈Ω

0 . xε ∈ Ω

We claim that mε → ∞ as ε → 0. Clearly this implies part (a). We prove this claim by an indirect argument and divide the proof into several steps. S TEP 1. If mε  M for some constant M and all ε > 0, then d(xε , ∂Ω0 ) → 0. Ω

Since λ
λ1 0 (φ), we must have uε L∞ (Ω) → ∞ as ε → 0, for otherwise uε increases to a positive solution of (2.1) with ε = 0 as ε decreases to 0, contradicting the fact that

Bifurcation and related topics

139

(2.2) is a necessary condition for (2.1) to possess a positive solution. Let us now pick up a sequence εn → 0, and define uˆ n = un /un ∞ , where un = uελn . We easily see that & % p−1 p −
uˆ n = λuˆ n − b(x) + εn un ∞ uˆ n ,

uˆ n |∂Ω = 0.

It follows that

−
uˆ n  λuˆ n .

(2.12)

Therefore 

Ω

2

|∇ uˆ n | dx  λ



Ω

uˆ 2n dx  λ|Ω|,

and {uˆ n } is bounded in W01,2 (Ω). This implies that, subject to a subsequence, uˆ n converges weakly in W01,2 (Ω) and strongly in Lq (Ω) (for all q > 1) to some uˆ ∈ W01,2 (Ω). We claim that uˆ ≡ 0. Indeed, if uˆ = 0, then from uˆ n → 0 in Lq (Ω) for all q > 1 we deduce (−
)−1 uˆ n → 0 in the C 1 norm; in particular, (−
)−1 uˆ n → 0 uniformly in Ω. From (2.12) we deduce 0  uˆ n  λ(−
)−1 uˆ n → 0, contradicting the fact that uˆ n ∞ = 1. This proves that uˆ ≡ 0. An application of Lemma 2.5 shows that un is bounded from above by U λ on Ω+ := 0 . From this we easily see that uˆ ≡ 0 on Ω+ . Thus, as ∂Ω0 is smooth, u| Ω\Ω ˆ Ω0 ∈ 1,2 W0 (Ω0 ). If un ∞ = u(xn ), xn ∈ Ω. Then by Bony’s maximum principle (see [Bo] and [L]), there exists a sequence x˜k → xn such that limk→∞
un (x˜k )  0 and hence, from the equation for un , we obtain % & 0  λun (xn ) − b(xn ) + εn un (xn )p . p−1

p−1

It follows that εn un ∞  λ. Hence we may assume that εn un ∞ → ξ for some ξ
0. Now we multiply the equation for uˆ n by an arbitrary ψ ∈ C0∞ (Ω0 ) and integrate over Ω0 , and pass to the limit n → ∞, to obtain that 

Ω0

∇ uˆ · ∇ψ dx =



Ω0

 λuˆ − ξ uˆ p ψ dx.

That is to say that u| ˆ Ω0 is a weak solution to

 −
u = λ − ξ up−1 u,

u|∂Ω0 = 0.

By the weak Harnack inequality, we deduce uˆ > 0 in Ω0 .

140

Y. Du

0 . By standard From the equation for uˆ n , we see that −
uˆ n is uniformly bounded on Ω interior Lp -theory for elliptic equations (see [GT]), we find that uˆ n is bounded in W 2,q (Ω ′ ) for any q > 1 and any compact subdomain Ω ′ of Ω0 . By the Sobolev embedding theorem ′ ). As uˆ > 0 on Ω0 , (see [GT]), we know that subject to a subsequence, uˆ n → uˆ in C 1 (Ω and un ∞ → ∞, we find that un (x) → ∞ uniformly on any compact subset of Ω0 . As ε → uελ is monotone, uελ → ∞ uniformly on any compact subset of Ω0 as ε → 0. Thus we must have d(xε , ∂Ω0 ) → 0 as ε → 0. S TEP 2. If mε < M for some M and all ε > 0, then {∂uελ (xε )/∂νε } is bounded from above, where νε is a unit vector in RN to be specified later. It suffices to show that, for any sequence εn → 0, {∂uελn (xεn )/∂νεn } has a subsequence which is bounded from above. Let us denote  un = uελn , xn = xεn and Ωn = x ∈ Ω0 : d(x, ∂Ω0 )
d(xn , ∂Ω0 ) .

Note that if xn ∈ ∂Ω0 , then Ωn = Ω0 , and if Ωn is different from Ω0 , then for large n, it is close to Ω0 by Step 1. Thus for any Ω ′ ⊂⊂ Ω0 , Ω ′ ⊂⊂ Ωn for all large n. Clearly un is an upper solution to the problem % & n , u|∂Ω = 0, u|∂Ωn = un (xn ), (2.13) −
u = λu − b(x) + 1 up in Ω \ Ω and 0 is a lower solution. Therefore (2.13) has a positive solution vn satisfying 0  vn  un in Ω \ Ωn . As un (xn ) = vn (xn ), it follows that ∂un (xn ) ∂vn (xn )  , ∂νn ∂νn

where νn is the unit normal vector of ∂Ωn at xn pointing inward of Ωn . Thus it suffices to show that ∂vn (xn )/∂νn is bounded. Clearly C0 := max{λ1/(p−1) , M} is an upper solution to (2.13). By Lemma 2.5, we conclude that vn  C0 . This implies that −
vn has an L∞ bound on Ω \ Ωn which is independent of n. Since, furthermore, (1) vn |∂Ωn is a constant which has a bound independent of n, and (2) for all large n, ∂Ωn is as smooth as Ω0 with the smoothness not depending on n, by the Lp -theory of elliptic equations up to the boundary (see, e.g., [GT]), we see that, for any q > 1, vn W 2,q (Ω\Ωn ) has a bound independent of n. By Sobolev embeddings and the uniform smoothness of Ωn , this implies that vn C 1 (Ω\Ω  n ) has a bound independent of n. In particular {|∇vn (xn )|} is bounded, and thus {∂vn (xn )/∂νn } is bounded, as required. S TEP 3. mε → ∞ as ε → 0. Otherwise we can find a sequence εn → 0 such that mεn is bounded. By Step 2, {∂un (xn )/∂νn } is bounded from above, where νn is the unit normal vector of ∂Ωn at xn pointing inward of Ωn . Here we follow the notation in Step 2. We show that this is impossible, and hence proving the claim. For all large n, ∂Ωn is as smooth as ∂Ω0 and hence it

Bifurcation and related topics

141

satisfies a uniform interior ball condition: There exists R > 0 such that for any large n and n and Bx ∩ ∂Ωn = {x}. x ∈ ∂Ωn , one can find a closed ball Bx of radius R such that Bx ⊂ Ω Let yn denote the center of Bxn and define 2

2

ψ(x) = e−σ |x−yn | − e−σ R , where σ is a positive number to be specified. We may assume that εn < 1 for all n. Then, −1/p and x ∈ Bxn \ B n , where B n = {x: |x − yn | < for any constant c satisfying 1 < c < εn R/2}, we have % & % & % &p
un (xn ) + cψ + λ un (xn ) + cψ − εn un (xn ) + cψ (p ' % 2 & −σ |x−yn |2 2 p un (xn ) +ψ
ce 4σ |x − yn | − 2N σ − εn c c  % &p 2
ce−σ R σ 2 R 2 − 2N σ − un (xn ) + ψ > 0,

if σ , c and n are large enough. We fix σ at such a value.  n Choose a compact set K ⊂⊂ Ω0 such that K ⊃ ∞ n=1 B . By the proof of Step 1, −1/p un → ∞ on K. Hence we can find a sequence cn → ∞ satisfying cn  εn and un (x)
M + cn ψ|∂B n

for all x ∈ ∂B n ⊂ K.

Thus, un is an upper solution to the problem


−
u = λu − εn up u|∂Bxn = un (xn ),

in Bxn \ B n , u|∂B n = un (xn ) + cn ψ|∂B n .

By our choice of σ , for all large n, un (xn ) + cn ψ is a lower solution to this problem. Using Lemma 2.5 we deduce un
un (xn ) + cn ψ in Bxn \ B n , and it follows that ∂un (xn ) ∂ψ(xn ) 2
cn = cn 2σ Re−σ R → ∞. ∂νn ∂νn This contradicts the conclusion in Step 2. Thus the claim and hence part (a) in conclusion (ii) of the theorem is proved. It remains to prove part (b). By the above proved part (a), we see that un |∂Ω0 → ∞ uniformly as n → ∞. By Lemma 2.5 we deduce un  un+1  U λ . Therefore un → u0  U λ as n → ∞. It follows that u0 is a positive solution of (2.3). Since U λ is the minimal positive solution, we must have u0 = U λ . The proof is complete. 

142

Y. Du

It was proved in [Du3], Part II, Proposition 2.3, that the minimal positive solution U λ 0 , varies continuously with λ ∈ R1 in the space X := C(Ω+ ∪ ∂Ω), where Ω+ := Ω \ Ω and X has the metric defined by d(u, v) =

∞ −n  2 dn (u, v) 1 + dn (u, v) n=1

with dn (u, v) = u − vC(Ωn ) ,


δ + : d(x, ∂Ω0 )
, Ωn := x ∈ Ω n

where δ > 0 is small enough. This implies that  Γ∞ := (λ, U λ ): λ ∈ R1

is a continuous curve in R1 × C(Ω+ ∪ ∂Ω). If we define  λ (x) = U




+∞, U λ (x),

0 , x∈Ω x ∈ Ω+ ,

 λ ): λ ∈ R1 } as a continuous bifurcation curve at infinity, then and consider Γ∞ = {(λ, U Theorems 2.4 and 2.6 can be interpreted as follows: (i) The positive solution curve Γ = {(λ, uλ )} of (2.1) bifurcates from the trivial solu tion curve Γ0 = {(λ, 0)} at λ = λΩ 0 , then joins the bifurcation curve from infinity Γ∞ at Ω0 λ = λ1 . (ii) As ε → 0, the positive solution curve Γε = {(λ, uελ )} of (2.10) approaches Γ when Ω0 Ω0  λ ∈ (λΩ 1 , λ1 ), and it approaches Γ∞ when λ
λ1 . ε In order to better understand the profile of uλ (which is the unique positive solution of (2.10)), we consider wλε := ε p−1 uελ . It is easily seen that wλε is the unique positive solution of the problem % & −
w = λw − 1 + ε −1 b(x) w p ,

w|∂Ω = 0.

(2.14)

Ω

0 ε If λ ∈ (λΩ 1 , λ1 ), then by Theorem 2.6(i), we see that wλ → 0 uniformly in Ω as ε → 0. Ω0 We now consider the case that λ > λ1 . If we denote by θλ the unique positive solution of

−
w = λw − w p ,

w|∂Ω = 0,

then by Lemma 2.5 we see that wλε  θλ . Also by Lemma 2.5, we find that wλε is nonincreasing with ε. Therefore wλ0 (x) = limε→0 wλε (x) ∈ [0, θλ (x)] exists. Furthermore, on any compact subset K of Ω0 , −
wλε = λwλε − (wλε )p has an L∞ bound from above independent of ε. By the Lp estimates and Sobolev embedding theorem we find that

Bifurcation and related topics

143

wλε converges to wλ0 in C 1 (K). By Theorem 2.6(ii)(b), we see that wλε → 0 uniformly \Ω 0 . \Ω 0 . It follows that w 0 = 0 over Ω on any compact subset of Ω λ 0 Let θλ denote the unique positive solution of −
w = λw − w p ,

w|∂Ω0 = 0.

(2.15)

By Lemma 2.5 we obtain wλε
θλ0 . Therefore wλ0
θλ0 in Ω0 . We show that wλ0 = θλ0 in Ω0 . Indeed, from the inequality −
wλε  λwλε  λθλ we deduce that {wλε } is bounded in W01,2 (Ω) and therefore by a compactness argument, wλε → wλ0 weakly in W01,2 (Ω) and strongly in Lq (Ω) for any q > 1 (we also use the fact that wλε ∞ is bounded). Since wλ0 = 0 in Ω+ and ∂Ω0 is smooth, we conclude that wλ0 |Ω0 ∈ W01,2 (Ω0 ). It follows easily that wλ0 |Ω0 is a week positive solution of (2.15). By standard elliptic regularity, wλ0 is also a classical positive solution. But θλ0 is the unique such solution. Therefore wλ0 = θλ0 in Ω0 .  and as ε → 0, w ε → w 0 in Lq (Ω), We now find that wλ0 is a continuous function in Ω, λ λ  \ ∂Ω0 . We claim that this and the convergence is uniform on any compact subset of Ω  From the above discussions, it is clear that we need only convergence is uniform over Ω. prove the following conclusion: For any given δ > 0, there exists σ0 > 0 so that wλε (x) < δ

 ∀ε ∈ (0, σ0 ), ∀x ∈ Sσ0 := x ∈ Ω: d(x, ∂Ω0 ) < σ0 .

(2.16)

Ω

σ Denote Ωσ = {x ∈ Ω: d(x, Ω0 ) < σ }. Since λ > λ1 0 , for all small σ > 0, λ > λΩ 1 . Therefore the problem

−
w = λw − w p ,

w|∂Ωσ = 0,

has a unique positive solution θλσ . If we extend θλσ to be 0 outside Ωσ , then a simple compactness argument shows that θλσ → θλ0 as σ → 0 uniformly in Ω. In particular, we σ1 and can find σ1 > 0 small so that λ > λΩ 1 θλσ1 (x)
0 on Ω \ Ωσ1 . Then by what has been proved above, we have w˜ λε → θλσ1 uniformly on Ωσ1 /2 , where w˜ λε is the unique positive solution to % & p ˜ −
w = λw − 1 + ε −1 b(x) w ,

w|∂Ω = 0.

By Lemma 2.5 we deduce wλε  w˜ λε . Choose σ0  σ1 /2 such that for ε < σ0 , w˜ λε  θλσ1 +

δ 2

on Ωσ1 /2 .

144

Y. Du

Therefore, by (2.17), for ε ∈ (0, σ0 ) and x ∈ Sσ0 ⊂ Sσ1 /2 , wλε  w˜ λε  θλσ1 +

δ < δ, 2

that is, (2.16) holds. We have thus proved the following theorem. Ω

T HEOREM 2.7. Suppose that λ > λ1 0 . Then the unique positive solution wλε of (2.14)  as ε → 0. converges uniformly to wλ0 on Ω Let us observe that Theorem 2.7 gives a clear description of the pattern of wλε for small ε > 0: It is close to 0 over Ω \ Ω0 and close to a definite positive function θλ0 over Ω0 . 2.3. Comments and related results Positive solutions of problem (2.1) seem first considered by Ouyang [Ou], motivated by some geometric questions. Soon after, the results in [Ou] were extended in several directions by a number of authors. For example, Fraile, Koch-Medina, López-Gómez and Merino [FKLM] used an upper and lower solution argument to obtain a priori bounds, which greatly simplified the arguments in [Ou]. The proof of our Lemma 2.1 follows the approach of [FKLM]. Theorem 2.4 was proved in [DH]. Under some extra conditions on b(x) near ∂Ω0 , similar results were proved in [GGLS]. More references can be found in [DH]. Theorem 2.6 is taken from [Du3], Part II, and Theorem 2.7 from [DL]. The restriction that Ω0 is connected can be relaxed to the situation that it has finitely many components, each with smooth boundary; the techniques here can be easily adapted to deal with this case. Related results can be found, for example, in [Lop2] and [DL]. 0 is not contained in Ω, then some of the techniques here collapse, though it is If Ω expected that similar results hold. The case that ∂Ω0 ∩ ∂Ω = ∅ was discussed in [DG]. For the existence and uniqueness of positive solutions of (2.1), the smoothness condition on b(x) and Ω0 can be greatly relaxed, see [dP] and [Da8]. How to extend Theorems 2.4 and 2.6 to these situations remains to be investigated, though partial results were obtained in [dP]. We now come back to (2.1). By Theorem 2.4, we know that the branch of positive Ω0 solutions bifurcating from the trivial solution at λ = λΩ 1 blows up as λ approaches λ1 . If λΩ k denotes the kth eigenvalue of −
under Dirichlet boundary conditions, and we Ω define λk 0 similarly, then by Rabinowitz’s global bifurcation theorem, a global branch of nontrivial solutions Γk of (2.1) bifurcates from the trivial solution branch at λ = λΩ k if Ω0 λΩ is of odd algebraic multiplicity. If Γ is unbounded, then must it blow up at λ = λ k k ? k From Section 2.1, we know that this is the case when k = 1. In [DO] several special cases were considered where bifurcation branches starting from λΩ k (k > 1) indeed blow up at Ω0 λ = λk . Whether this is true in general remains open. Nevertheless, we have the following Ω result (see [DO]), which shows that the set {λk 0 : k
1} contains all the possible λ values where bifurcation from infinity can occur.

Bifurcation and related topics

145

T HEOREM 2.8. Given any large positive constant Λ and open neighborhood V of the finite set  Ω Ω M = λk 0 : λk 0  Λ, k
1 ,

we can find a constant C depending on Λ and V such that any solution (λ, u) of (2.1) with λ  Λ and λ ∈ / V satisfies uL∞  C. P ROOF. We argue indirectly. Suppose that we can find a sequence of solutions {(λn , un )} of (2.1) such that λn  Λ, λn ∈ / V , and un L∞ → ∞. Clearly we have 

Ω

|∇un | dx  λn



Ω

u2n dx.

It follows that λn
λΩ 1 . Therefore, by passing to a subsequence, we may assume that , Λ] \V. λn → λˆ and λˆ ∈ [λΩ 1 0 . C LAIM 1. {un } is uniformly bounded on any compact subset of Ω+ := Ω \ Ω To prove Claim 1, we let K be an arbitrary compact subset of Ω+ . By our assumption on b, for some small neighborhood U of K, there exists τ > 0 such that b(x)
τ on U . Denote by Vλ the unique positive solution of −
u = λu − τ up ,

u|∂U = ∞,

whose existence and uniqueness is well known (see [MV]). Choose λ∗ such that λ∗ > λn for all n
1. We want to show that un  Vλ∗ on K. Otherwise, we can find some n
1 and a domain U0 ⊂⊂ U such that un > Vλ∗ in U0 and un = Vλ∗ on ∂U0 . Hence on U0 , we have p

−
un = λn un − b(x)un and

p

−
Vλ∗
λ∗ Vλ∗ − b(x)Vλ∗ . An application of Lemma 2.5 on U0 yields un  Vλ∗ on U0 , a contradiction. Similarly, we find −un  Vλ∗ on U . Therefore |un |  Vλ∗ and Claim 1 follows.

∀x ∈ U,

146

Y. Du

C LAIM 2. λˆ ∈ M. Clearly this contradicts the fact that λˆ ∈ / V . Therefore our proof is complete once Claim 2 is proved. Denoting uˆ n = un /un L∞ we find that 

Ω

2

|∇ uˆ n | dx  λn



Ω

uˆ 2n dx  Λ|Ω|.

Thus {uˆ n } is a bounded sequence in W01,2 (Ω). It follows that, subject to a subsequence, uˆ n converges to some uˆ weakly in W01,2 (Ω) and strongly in L2 (Ω). As uˆ n has L∞ norm 1, we find that uˆ n → uˆ in Lq (Ω) for any q > 1. By Claim 1, we see that uˆ ≡ 0 over Ω+ . Since ˆ Ω0 ∈ W01,2 (Ω0 ). ∂Ω0 is smooth, it is well known that this implies v0 := u| We show next that v0 ≡ 0. Otherwise, uˆ ≡ 0 over Ω and hence uˆ n → 0 in Lq for all q > 1. By Kato’s inequality (see [Kato]), we have, in the weak sense, −
|uˆ n |  −

uˆ n
uˆ n  λn |uˆ n |  Λ|uˆ n |. |uˆ n |

Therefore 0  |uˆ n |  Λ(−
)−1 |uˆ n |. By standard elliptic regularity we find (−
)−1 |uˆ n | → 0 uniformly in Ω. It follows that uˆ n L∞ → 0 as n → ∞. But this contradicts the fact that uˆ n L∞ = 1. Hence we have proved that v0 ≡ 0. We now multiply the equation satisfied by uˆ n by an arbitrary φ ∈ C0∞ (Ω0 ), integrate by parts and find 

Ω0

∇ uˆ n · ∇φ dx =

Letting n → ∞ we obtain 

Ω0

∇v0 · ∇φ dx =



λn uˆ n φ dx.



λˆ v0 φ dx.

Ω0

Ω0

This implies that v0 ∈ W01,2 (Ω0 ) solves ˆ −
v = λv,

v|∂Ω0 = 0,

(2.18)

in the weak sense. Standard elliptic regularity shows that v0 is also a classical solution Ω of (2.18). As we have already proved that v0 ≡ 0, we must have λˆ = λk 0 for some k
1. Thus λˆ ∈ M. This proves Claim 2 and hence concludes the proof of Theorem 2.8. 

Bifurcation and related topics

147

3. Bifurcation and monotonicity: A heterogeneous competition system Bifurcation and monotonicity have been combined to produce many nice results in nonlinear analysis, and a collection of these techniques and results may be found in [Am]. In this section we present some new applications. It is in general not easy to capture the influence of heterogeneous spatial environment on population models. Traditionally population models were considered in homogeneous spatial environment, and hence all the coefficients appearing in the models are chosen to be positive constants. To include spatial variations of the environment, these constant coefficients should be replaced by positive functions of the space variable x. However, the mathematical techniques developed to study these models are ironically either not sensitive to this change, in which case the effects of heterogeneous spatial environment are difficult to observe in the mathematical analysis, or the techniques are too sensitive to this change and become inapplicable when the constant coefficients are replaced by functions. In this section we use a competition model to demonstrate that bifurcation techniques are useful in capturing these spatial effects. Here we combine the bifurcation arguments with a certain monotonicity property of the system. A key in this approach is the following observation: The behavior of the model is very sensitive to certain coefficient functions becoming small in part of the underlying spatial region. To make our ideas more transparent, we consider the following simplified steady-state competition system  % & 2   −
u = λu − b(x) + ε u − cuv, (3.1) −
v = µv − v 2 − duv,   u|∂Ω = v|∂Ω = 0.

Here Ω is a bounded smooth domain in RN (N
2) and λ, µ, c, d are constants, with c > 0, d > 0. The function b(x) is as in (2.1). We will see that when ε is small, the profiles of certain stable positive solutions of (3.1) are determined by the behavior of b(x). Our approach makes use of the facts that the global bifurcation branches for (3.1) with ε = 0 differs considerably to that for (3.1) with any ε > 0. By carefully following the changes of the global bifurcation branches of (3.1) as ε shrinks to 0, we will be able to obtain a rather detailed global picture of the solution branches and to describe the patterns of the stable solutions. The techniques here can be used to study systems much more general than (3.1). We refer to [Du3,Du4] for discussions of the background.

3.1. Global bifurcation We will fix ε > 0 and apply a global bifurcation analysis to (3.1), using µ as a bifurcation parameter. We are only interested in nonnegative solutions. Firstly let us observe some preliminary results. Clearly (u, v) = (0, 0) is always a solution to (3.1), which is called the trivial solution. If (u, 0) = (0, 0) is a nonnegative solution of (3.1), then u is a positive solution of (2.10) with p = 2, namely % & −
u = λu − b(x) + ε u2 ,

u|∂Ω = 0.

(3.2)

148

Y. Du

It is well known that (3.2) has a unique positive solution when λ > λΩ 1 , and there is no Ω , we denote the unique positive solution of (3.2) positive solution when λ  λΩ . For λ > λ 1 1 by φλε , and when the ε dependence is not emphasized, we will simply denote it by φλ . If (0, v) = (0, 0) is a nonnegative solution of (3.1), then v is a positive solution of −
v = µv − v 2 ,

v|∂Ω = 0.

(3.3)

Similarly to (3.2), there is a unique positive solution θµ to (3.3) when µ > λΩ 1 , and no . The solutions (φ , 0) and (0, θ ) are called semitrivial positive solution exists if µ  λΩ λ µ 1 solutions of (3.1). We next discuss the positive solutions of (3.1), where both components u and v are positive in Ω. (By the strong maximum principle, if (u, v) solves (3.1) and u
, ≡ 0, v
, ≡ 0, then u > 0, v > 0 in Ω.) Suppose that (3.1) has a positive solution. Then from the first equation we obtain

% &  Ω b(x) + ε u + cv > λΩ λ = λΩ 1 1 (0) = λ1 .

Similarly, from the second equation we deduce µ > λΩ 1 . Therefore the following is a necessary condition for (3.1) to possess a positive solution: λ > λΩ 1 ,

µ > λΩ 1 .

(3.4)

So from now on we fix λ > λΩ 1 . We also assume that ε, c and d are fixed; µ will be considered as our bifurcation parameter. We now obtain some rough estimates for positive solutions of (3.1). Let (u, v) be a positive solution of (3.1). Then u is a lower solution of (3.2). Hence we can apply Lemma 2.5 to deduce that u  φλ . Similarly, v  θµ . Using the above estimate for u we deduce −
v + dφλ v
µv − v 2 , and by Lemma 2.5, v
vµ , where vµ is the unique positive solution of −
v + dφλ v = µv − v 2 ,

v|∂Ω = 0,

provided that µ > λΩ ˆ > 0 large enough so that 1 (dφλ ). We now show that there exists µ (3.1) has no positive solution when µ > µ. ˆ Indeed, we have

% &  b(x) + ε u + cv
λΩ λ = λΩ 1 (cvµ ). 1

(3.5)

Since vµ → ∞ as µ → ∞ uniformly on any compact subsets of Ω, we can show that ˆ for some large µ. ˆ λΩ 1 (cvµ ) → ∞ as µ → ∞. Hence (3.5) implies that µ  µ To summarize, we have the following result.

Bifurcation and related topics

149

T HEOREM 3.1. Suppose λ > λΩ 1 and ε, c, d are fixed positive constants. Then there exists µˆ > 0 such that if (3.1) has a positive solution (u, v), then λΩ ˆ 1 < µ < µ,

u  φλ ,

v  θµ  θµˆ .

We now transform (3.1) into an abstract equation and apply bifurcation and monotonicity arguments to study its positive solution set. Choose M > 0 large enough such that for µ ∈ [0, 1 + µ] ˆ and 0  u  ξ := 1 + φλ ∞ , 0  v  η := 1 + θµˆ ∞ , % & g(u, v) := Mu + λu − b(x) + ε u2 − cuv

is strictly increasing in u, and

h(µ, u, v) := Mv + µv − v 2 − duv is strictly increasing in v. Then define

 A(µ, u, v) = (−
+ M)−1 g(u, v), h(µ, u, v) ,

(u, v) ∈ E0 , µ ∈ R1 ,

 × C 1 (Ω):  u|∂Ω = v|∂Ω = 0}. where E0 := {(u, v) ∈ C 1 (Ω) Clearly P0 := {(u, v) ∈ E0 : u  0, v
0} is a cone in E0 . It introduces a partial ordering in E0 : (u1 , v1 ) P0 (u2 , v2 )

if and only if (u2 − u1 , v2 − v1 ) ∈ P0 .

Denote J := [0, 1 + µ], ˆ

 A0 := (u, v) ∈ E0 : 0  u < ξ, 0  v < η in Ω .

It is easily seen, by the positivity of (−
+ M)−1 , that for fixed µ ∈ J , A(µ, u, v) is increasing over A0 in the order P0 , namely (u1 , v1 ), (u2 , v2 ) ∈ A0 ,

(u1 , v1 ) P0 (u2 , v2 )

implies

A(µ, u1 , v1 ) P0 A(µ, u2 , v2 ).

Moreover, for fixed (u, v) ∈ A0 , A(µ, u, v) is increasing in µ: µ1  µ2

implies A(µ1 , u, v) P0 A(µ2 , u, v).

If we use (u1 , v1 ) ≪P0 (u2 , v2 ) to mean (u2 − u1 , v2 − v1 ) ∈ int P0 and use (u1 , v1 ) λ1 (cθµ ), and hence the problem % & −
u = λu − b(x) + ε u2 − cθµ u, u|∂Ω = 0

has a unique positive solution u0 . If we denote v0 := θµ , then it is easily checked that A(µ, u0 , v0 ) λu˜ − b(x) + ε u˜ 2 − cθµ u.

We now apply Lemma 2.5 and conclude that u˜
u0 in Ω. Therefore (u, ˜ v) ˜ P0 (u, ˜ v). ˜ If we define (un , vn ) := A(µ, un−1 , vn−1 ),

n = 1, 2, . . . ,

we easily deduce that (u, ˜ v) ˜ P0 (un , vn ) P0 (un−1 , vn−1 ) P0 (u0 , v0 ),

n = 2, 3, . . . .

(3.12)

A standard compactness argument shows that (un , vn ) → (uµ , v µ ) in E0 and (uµ , v µ ) = A(µ, uµ , v µ ). By (3.12), we know that (uµ , v µ ) is a positive solution of (3.1). We claim that it is the maximal positive solution. Indeed, if (u, v) is any positive solution of (3.1), then by Theorem 3.1, v  θµ and hence % & −
u
λu − b(x) + ε u2 − cθµ u.

By Lemma 2.5, this implies that u
u0 . Therefore (u, v) P0 (u0 , v0 ). It follows from this inequality and the monotonicity of A that (u, v) P0 (un , vn ) for all n
0. Thus (u, v) P0 (uµ , v µ ).  L EMMA 3.4. If µ∗ > µ0 , then for every µ ∈ (µ0 , µ∗ ), (3.1) has a minimal positive solution (uµ , vµ ). P ROOF. Let µ ∈ (µ0 , µ∗ ). Then µ > λΩ 1 (dφλ ) and hence the problem −
v = µv − v 2 − dφλ v,

v|∂Ω = 0,

152

Y. Du

has a unique positive solution v0 . Denote u0 = φλ . Then (u0 , v0 ) satisfies A(µ, u0 , v0 )
P0 (u0 , v0 ). Choose µ˜ ∈ (µ, µ∗ ] such that (3.1) with µ = µ˜ has a positive solution (u, ˜ v). ˜ Then it is easily checked that A(µ, u, ˜ v) ˜ µ0 , then (µ, uµ , vµ ) ∈ Γ for all µ ∈ (µ0 , µ∗ ). P ROOF. Clearly µ0
inf{µ: (µ, u, v) ∈ Γ } and µ0  sup{µ: (µ, u, v) ∈ Γ }. Therefore we have nothing to prove if µ∗ = µ0 and µ∗ = µ0 . Suppose µ∗ < µ0 . Let us define, for µ ∈ (µ∗ , µ0 ),

where

%   ∆µ := [µ, ∞) × uµ , v µ , ∞ ,

% µ µ   

 u , v , ∞ := (u, v) ∈ E0 : (u, v)
P0 uµ , v µ .

/ ∆µ , it follows from the conSince Γ connects (µ0 , 0, θµ0 ) ∈ int ∆µ and (µ0 , φλ , 0) ∈ µ nectedness of Γ that Γ ∩ ∂∆ = ∅. Let (µ, ˜ u, ˜ v) ˜ ∈ Γ ∩ ∂∆µ . We claim that (µ, ˜ u, ˜ v) ˜ = µ µ (µ, u , v ). Clearly  & % %  & % % ∂∆µ = {µ} × uµ , v µ , ∞ ∪ [µ, ∞) × ∂ uµ , v µ , ∞ .

If µ˜ = µ, then since (uµ , v µ ) is the maximal positive solution of (3.1), we necessarily have (u, ˜ v) ˜ P0 (uµ , v µ ). On the other hand, we have (u, ˜ v) ˜ ∈ [(uµ , v µ ), ∞). Therefore µ µ (u, ˜ v) ˜ = (u , v ). We show next that µ˜ > µ is impossible. Indeed, if µ˜ > µ, then we must have (u, ˜ v) ˜ ∈ ∂[(uµ , v µ ), ∞), i.e.,

 (u, ˜ v) ˜
P0 uµ , v µ ,

 (u, ˜ v) ˜ ≫P0 uµ , v µ .

Bifurcation and related topics

153

By the monotonicity properties of the operator A, we have

  (u, ˜ v) ˜ = A(µ, ˜ u, ˜ v) ˜ >P0 A(µ, u, ˜ v) ˜
P0 A µ, uµ , v µ = uµ , v µ .

˜ v) ˜ ≫P0 A(µ, uµ , v µ ) and Now from (u, ˜ v) ˜ >P0 (uµ , v µ ), we further obtain A(µ, u, µ µ µ hence (u, ˜ v) ˜ ≫P0 (u , v ). Therefore (µ, ˜ u, ˜ v) ˜ ∈ int ∆ , contradicting the assumption that (µ, ˜ u, ˜ v) ˜ ∈ ∂∆µ . This proves assertion (i). The proof for (ii) is similar, where we use

& ∆µ = (−∞, µ] × −∞, (uµ , vµ )

with

& 

−∞, (uµ , vµ ) := (u, v) ∈ E0 : (u, v) P0 (uµ , vµ ) ,

and the fact that Γ ∩ ∂∆µ = {(µ, uµ , vµ )}. We omit the details. Clearly, (3.11) is a consequence of (i) and (ii).



T HEOREM 3.6. (i) If µ∗ < min{µ0 , µ0 }, then (3.1) has a maximal positive solution (uµ∗ , v µ∗ ) for µ = µ∗ and it has at least two positive solutions for each µ ∈ (µ∗ , min{µ0 , µ0 }). Moreover, all these solutions can be chosen from Γ . (ii) If µ∗ > max{µ0 , µ0 }, then (3.1) has a minimal positive solution for µ = µ∗ and it has at least two positive solutions for each µ ∈ (max{µ0 , µ0 }, µ∗ ). Moreover, all these solutions can be chosen from Γ . P ROOF. Suppose µ∗ < min{µ0 , µ0 }. Let {µn } ⊂ (µ∗ , µ0 ) be a decreasing sequence converging to µ∗ . By Theorem 3.5, (3.1) with µ = µn has a maximal positive solution (un , vn ) := (uµn , v µn ). Using the estimates in Theorem 3.1 and a compactness argument we easily see that, subject to a subsequence, (un , vn ) → (u∗ , v∗ ) in E0 , and (u∗ , v∗ ) is a nonnegative solution of (3.1) with µ = µ∗ . Since µn is decreasing, by the monotonicity property of A we easily deduce that (un , vn ) ≫P0 (un+1 , vn+1 ). Therefore u∗
un > 0 in Ω. We claim that v∗ ≡ 0 in Ω. Indeed, if v∗ ≡ 0, then from the equation for un in (3.1) we easily deduce that un → φλ in E0 , and hence u∗ = φλ . But then from the equation for vn we deduce Ω µn = λΩ 1 (vn + dun ) → λ1 (dφλ ) = µ0 ,

a contradiction to our assumption that µn → µ∗ . Therefore (u∗ , v∗ ) is a positive solution of (3.1) with µ = µ∗ . If (u, v) is any positive solution of (3.1) with µ = µ∗ , then by making use of the monotonicity property of A we deduce (u, v) ≪P0 (un , vn ) for all n
1. It follows that (u, v) P0 (u∗ , v∗ ). Therefore (u∗ , v∗ ) is the maximal solution. The proof of Theorem 3.5 shows that (µ∗ , u∗ , v∗ ) ∈ Γ . To show that (3.1) has at least two positive solutions for µ ∈ (µ∗ , min{µ0 , µ0 }), we use the following simple result from general point set theory, whose proof can be found in [DB].

154

Y. Du

L EMMA 3.7. Suppose that X is a Banach space, C is a connected set in X and O an open set in X such that ∂O ∩ C consists of a single point. Then C \ O is a connected set. Now for fixed µ ∈ (µ∗ , min{µ0 , µ0 }), we take C = Γ and O = int ∆µ˜ , where µ˜ ∈ (µ∗ , µ). By the proof of Theorem 3.5, we know that ∂∆µ˜ ∩ Γ = {(µ, ˜ uµ˜ , v µ˜ )}. Hence by µ ˜ Lemma 3.7, the set Γ \ int ∆ is connected. Since (µ∗ , u∗ , v∗ ) ∈ Γ \ ∆µ˜ and (µ0 , φλ , 0) ∈ Γ \ ∆µ˜ , by the connectedness of Γ \ ∆µ˜ and the fact that µ ∈ (µ∗ , µ0 ), we can find a point (µ, u, v) ∈ Γ \ ∆µ˜ . Hence (u, v) is a positive solution of (3.1). Since µ˜ < µ, we have (µ, uµ , v µ ) ∈ int ∆µ˜ . Therefore (u, v) and (uµ , v µ ) are different positive solutions of (3.1), and both (µ, u, v) and (µ, uµ , v µ ) belong to Γ . This proves (i). The proof of (ii) is parallel and we omit the details.  R EMARK 3.8. By the connectedness of Γ , if µ∗ < µ∗ , then for any µ ∈ (µ∗ , µ∗ ), (3.1) has a positive solution on Γ . We will show in Section 3.3 that µ∗ > max{µ0 , µ0 } if ε > 0 is small enough.

3.2. Stability analysis Suppose that (u0 , v0 ) is a positive solution of (3.1). We want to know whether it is stable when considered as a steady-state of the corresponding parabolic problem  % & 2   ut −
u = λu − b(x) + ε u − cuv, x ∈ Ω, t > 0, x ∈ Ω, t > 0, vt −
v = µv − v 2 − duv,   u = v = 0, x ∈ ∂Ω, t > 0.

(3.13)

T HEOREM 3.9. Suppose that (u0 , v0 ) is a positive solution of (3.1), and there exists (h, k) ∈ P0 \ {(0, 0)} and σ ∈ R1 such that  % &   −
h = λh − 2 b(x) + ε u0 h − cv0 h − cu0 k + σ h, −
k = µk − 2v0 k − du0 k − dv0 h + σ k,   h|∂Ω = k|∂Ω = 0.

(3.14)

Then (u0 , v0 ) is asymptotically stable in E0 if σ > 0, and it is unstable if σ < 0. P ROOF. Since u0 > 0 and v0 > 0 in Ω, we easily see from (3.14) that h ≡ 0 and k ≡ 0. Moreover, from the first equation in (3.14), we obtain −
h  C(x)h in Ω for some  Therefore by the maximum principle and Hopf boundary lemma, we deduce C ∈ C(Ω). h < 0 in Ω and ∂ν h > 0 on ∂Ω. We similarly deduce k > 0 in Ω and ∂µ k < 0 on ∂Ω. This implies that (h, k) ∈ int P0 .

Bifurcation and related topics

155

For δ ∈ R1 let us denote uδ := u0 + δh, vδ := v0 + δk. Then a simple calculation gives % &  −
uδ = λuδ − b(x) + ε u2δ − cuδ vδ  

% &    + σ + δ b(x) + δ + cδk δh,  −
vδ = µvδ − vδ2 − duδ vδ + (σ + δ + dδh)δk,    uδ |∂Ω = vδ |∂Ω = 0.

(3.15)

Due to (3.15) we can find δ0 = δ0 (σ ) > 0 small enough so that, if σ > 0, then for all δ ∈ (0, δ0 ],  % & 2   −
uδ < λuδ − b(x) + ε uδ − cuδ vδ , 2

−
vδ > µvδ − vδ − duδ vδ ,   uδ |∂Ω = vδ |∂Ω = 0,

and for all δ ∈ [−δ0 , 0),  % & 2   −
uδ > λuδ − b(x) + ε uδ − cuδ vδ , 2

−
vδ < µvδ − vδ − duδ vδ ,   uδ |∂Ω = vδ |∂Ω = 0.

Hence (uδ0 , vδ0 ) is an upper solution of (3.1) and (u−δ0 , v−δ0 ) is a lower solution of (3.1). It is well known (see [Sm]) that the semiflow generated by the solution of (3.13) preserves the order P0 , and by the theory of monotone dynamical systems (see [Ma] or [Hir]), the following hold: (i) the unique solution ( u(x, t), v(x, t)) of (3.13) with initial data (u−δ0 , v−δ0 ) increases in the order P0 as t increases; (ii) the unique solution (u(x, ¯ t), v(x, ¯ t)) of (3.13) with initial data (uδ0 , vδ0 ) decreases in the order P0 as t increases; (iii)

 

¯ t), v(·, ¯ t) P0 (uδ0 , vδ0 ) (u−δ0 , v−δ0 ) P0 u(·, t), v(·, t) P0 u(·,

∀t > 0;

(iv) ( u(x), v(x)) = limt→∞ ( u(x, t), v(x, t)) and (u(x), ¯ v(x)) ¯ = limt→∞ (u(x, ¯ t), v(x, ¯ t)) exist and they are solutions of (3.1). Therefore we have ¯ v) ¯ P0 (uδ0 , vδ0 ). (u−δ0 , v−δ0 ) P0 ( u, v ) P0 (u, Define  δ∗ = inf δ ∈ [0, δ0 ]: ( u, v )
P0 (u−δ , v−δ ) .

Then 0  δ∗  δ0 . If δ∗ > 0, then

( u, v ) = A(µ, u, v )
P0 A(µ, u−δ∗ , v−δ∗ ) ≫P0 (u−δ∗ , v−δ∗ ).

156

Y. Du

Therefore for all δ < δ∗ but close to δ∗ , we also have ( u, v )
P0 (u−δ , v−δ ), contradicting the definition of δ∗ . This proves that ( u, v )
P0 (u0 , v0 ). Similarly we can show that (u, ¯ v) ¯ P0 (u0 , v0 ). Therefore we must have

   u(x), v(x) = u(x), ¯ v(x) ¯ = u0 (x), v0 (x) .

By the order preserving property of (3.13), we deduce that any solution (u(x, t), v(x, t)) of (3.13) with initial data taken from

satisfies

& % (u−δ0 , v−δ0 ), (uδ0 , vδ0 )  := (u, v) ∈ E0 : (u−δ0 , v−δ0 ) P0 (u, v) P0 (uδ0 , vδ0 )



  ¯ t), v(·, ¯ t) u(·, t), v(·, t) P0 u(·, t), v(·, t) P0 u(·,

∀t > 0.

It follows that (u(x, t), v(x, t)) → (u0 (x), v0 (x)) as t → ∞. By standard regularity for parabolic equations this convergence can be taken in the norm of E0 . This proves the asymptotic stability of (u0 , v0 ) in E0 , since [(u−δ0 , v−δ0 ), (uδ0 , vδ0 )] is an open neighborhood of (u0 , v0 ) due to (h, k) ∈ int P0 . If σ < 0, then for all small δ > 0, (uδ , vδ ) is a lower solution of (3.1). It follows that the unique solution (u, v) of (3.13) with initial date (uδ , vδ ) increases in P0 as t increases. By a simple comparison argument one sees that (u, v) stays bounded in the L∞ norm for all t > 0. Hence by the theory of monotone dynamical systems (u, v) converges to a steadystate of (3.13) as t → ∞, say (u∗ , v∗ ). Clearly (u∗ , v∗ )
P0 (uδ , vδ ) ≫P0 (u0 , v0 ). We can show that (u∗ , v∗ ) >P0 (uδ0 , vδ0 ) by a sweeping argument. Indeed, if we define  δ ∗ := sup η ∈ [δ, δ0 ]: (u∗ , v∗ )
P0 (uη , vη ) ,

then by the monotonicity property of A and the continuous dependence on δ of the (strict) lower solutions (uδ , vδ ), we easily deduce that δ ∗ = δ0 and (u∗ , v∗ ) >P0 (uδ0 , vδ0 ). There/ [(u−δ0 , v−δ0 ), (uδ0 , vδ0 )]. This implies that (u0 , v0 ) is unstable.  fore (u∗ , v∗ ) ∈ We now relate (3.14) to the spectral radius r(L) of the linear operator L : E0 → E0 , where L denotes the Frechet derivative of A(µ, u, v) with respect to (u, v) at (u0 , v0 ), namely L := D(u,v) A(µ, u0 , v0 ). T HEOREM 3.10. Let (u0 , v0 ) be a positive solution of (3.1). Then (3.14) has a solution (h, k) ∈ P0 \ {(0, 0)} with σ > 0 if r(L) < 1; it has a solution (h, k) ∈ P0 \ {(0, 0)} with σ < 0 if r(L) > 1. In the proof of Theorem 3.10, we will need the following version of the well-known Krein–Rutman theorem (see [De], Theorem 19.3, and [Am], Theorem 3.2).

Bifurcation and related topics

157

T HEOREM 3.11. Suppose that X is a Banach space with a positive cone P which has nonempty interior, and B is a compact linear operator in X, which is strongly positive: B(P \ {0}) ⊂ int P . Then r(B) > 0 and there exists a unique x0 ∈ int P such that Bx0 = r(B)x0 , x0  = 1; there exists φ ∈ X ∗ such that φ(x) > 0 for x ∈ P \ {0} and B ∗ φ = r(B)φ. Hence Bx − r(B)x ∈ / P \ {0} for any x ∈ X. P ROOF OF T HEOREM 3.10. It is easily checked that L is compact and strongly positive in E0 . Therefore we can apply Theorem 3.11 to find (h0 , k0 ) ∈ int P0 such that L(h0 , k0 ) = r(L)(h0 , k0 ). Suppose r(L) < 1 and define, for σ
0, Lσ (h, k) = L(h, k) + σ (−
+ M)−1 (h, k). Then Lσ is compact and strongly positive in E0 . Hence, by Theorem 3.11, r(Lσ ) > 0 and there exists a unique (hσ , kσ ) ∈ int P0 such that Lσ (hσ , kσ ) = r(Lσ )(hσ , kσ ),

  (hσ , kσ )

E0

= 1.

By the uniqueness of (hσ , kσ ) and a standard compactness argument, we easily see that r(Lσ ) varies continuously with σ . We also have 



σ −1 r(Lσ ) = r σ −1 L + (−
+ M)−1 → r (−
+ M)−1 > 0 as σ → ∞.

Therefore r(Lσ ) > 1 for all large σ . Since r(L0 ) = r(L) < 1, there exists σ0 > 0 such that r(Lσ0 ) = 1, i.e., Lσ0 (hσ0 , kσ0 ) = (hσ0 , kσ0 ). It is easily seen that this implies that (hσ0 , kσ0 ) solves (3.14) with σ = σ0 . Suppose now r(L) > 1. We then consider, for σ
1, the family of operators Lσ (h, k) := (−
+ σ M)−1

 × gu (u0 , v0 )h + gv (u0 , v0 )k, hu (u0 , v0 )h + hv (u0 , v0 )k .

We easily see that Lσ is a compact and strongly positive operator in E0 , and hence by Theorem 3.11, there exists a unique (hσ , k σ ) ∈ int P0 such that



  Lσ hσ , k σ = r Lσ hσ , k σ ,

 σ σ   h ,k 

E0

= 1.

As before, by the uniqueness of (hσ , k σ ) and a standard compactness argument, we easily see that r(Lσ ) varies continuously with σ . We claim that r(Lσ ) → 0 as σ → ∞. Indeed, if we write Lσ (hσ , k σ ) = r(Lσ )(hσ , k σ ) in its differential equation form, we obtain  &

σ −1 % σ σ  gu (u0 , v0 )hσ + gv (u0 , v0 )k σ ,   −
h + σ Mh = r L &

−1 % −
k σ + σ Mk σ = r Lσ hu (u0 , v0 )hσ + hv (u0 , v0 )k σ ,     σ  h = k σ  = 0. ∂Ω

∂Ω

158

Y. Du

Therefore there exists a large constant C > 0 such that     

 σ 2  σ 2 



 σ 2 h  + k  dx, ∇h  + σ M hσ 2 dx  r Lσ −1 C Ω



Ω

  



 σ 2 ∇k  + σ M k σ 2 dx  r Lσ −1 C

It follows that

 2C −1 σ → 0 as σ → ∞. r Lσ  M



Ω

Ω

 σ 2  σ 2  h  + k  dx.

0

0

0

Since r(L1 ) = r(L) > 1, we can find σ 0 > 1 such that r(Lσ ) = 1 and hence Lσ (hσ , 0 0 0 0 0 k σ ) = (hσ , k σ ). Writing this in the form of differential equations, we find that (hσ , k σ ) 0  solves (3.14) with σ = (1 − σ )M < 0. We now discuss the stability of the maximal positive solutions (uµ , v µ ) and the minimal positive solutions (uµ , vµ ) of (3.1). Suppose µ∗ < µ0 , and for µ ∈ (µ∗ , µ0 ), denote 

Lµ := D(u,v) A µ, uµ , v µ ,

  

O ∗ := µ ∈ µ∗ , µ0 : r Lµ < 1 .

By Theorems 3.9 and 3.10, we know that (uµ , v µ ) is asymptotically stable when µ ∈ O ∗ . Furthermore, by the implicity function theorem, O ∗ is an open set and (uµ , v µ ) varies continuously with µ for µ ∈ O ∗ . The following result shows that the measure of O ∗ is the same as that of the interval (µ∗ , µ0 ). T HEOREM 3.12. The set (µ∗ , µ0 ) \ O ∗ has measure zero in R1 . P ROOF. We use ideas from [Du1], Section 3. To simplify notation, we denote w(µ) := (uµ , v µ ). From the monotonicity of A, we easily deduce that w(µ1 ) ≪P0 w(µ2 ) when µ1 < µ2 . Moreover, w(µ) is right-continuous in µ. Indeed, suppose µn decreases to µ; by a compactness consideration we find that subject to a subsequence, w(µn ) → w in E0 and w = A(µ, w). Since w(µ) ≪P0 w(µn ), we deduce w(µ) P0 w. But w(µ) is the maximal solution and hence we necessarily have w = w(µ). This implies that limµ′ →µ+0 w(µ′ ) = w(µ). We divide the rest of the proof into several steps. S TEP 1. If w(µ) is discontinuous at µˆ ∈ (µ∗ , µ0 ), then w(µˆ − ) := limµ→µ−0 w(µ) exists ˆ ˆ and A(µ, ˆ w(µˆ − )) = w(µˆ − ) ≪P0 w(µ). Let {µn } ⊂ (µ∗ , µ0 ) be an arbitrary sequence increasing to µ. ˆ Then uµn (x)
uµn+1 (x)
uµˆ (x),

v µn (x)  v µn+1 (x)  v µˆ (x).

Therefore (u(x), ˆ v(x)) ˆ := limn→∞ (uµn (x), v µn (x)) exists and the limit is independent of ˆ := (u, ˆ v) ˆ the choice of {µn }. A simple compactness argument shows that w(µn ) → w

Bifurcation and related topics

159

ˆ = A(µ, ˆ We must have w ˆ = w(µ) in E0 . Clearly we have w ˆ w). ˆ for otherwise w(µ) would be continuous at µ. ˆ Since w(µ) ˆ is the maximal positive solution, we must have ˆ ≪P0 w(µ). ˆ P0 0, the above identities imply that β > 0 and zn = β −1 z0 + Lµˆ zn + o(1). From this and the compactness of Lµˆ , it follows that, subject to a subsequence, zn → zˆ in E0 , and zˆ = β −1 z0 + Lµˆ zˆ . Since zn >P0 0 and zn  = 1, we have zˆ >P0 0. If r(Lµˆ )
1, then we deduce

 Lµˆ (−ˆz) − r Lµˆ (−ˆz)
P0 zˆ − Lµˆ zˆ = β −1 z0 >P0 0,

a contradiction to the last conclusion of Theorem 3.11. Therefore we must have r(Lµˆ ) < 1. This proves Step 2. S TEP 3. If w(µ) is continuous at µˆ ∈ (µ∗ , µ0 ) and limµ→µ−0 (w(µ) ˆ − w(µ))/(µˆ − µ) = ˆ +∞, then there exists α(µ) > 0 such that w(µ) ˆ − w(µ)
α(µ)w0 , µˆ − µ

lim α(µ) = +∞,

µ→µ−0 ˆ

where w0 ∈ int P0 is the unique solution to Lµˆ w0 = r(Lµˆ )w0 , w0  = 1.

(3.16)

160

Y. Du

Define z(µ) := (w(µ) ˆ − w(µ))/w(µ) ˆ − w(µ). Then for µ → µˆ − 0, we have z(µ) = =

A(µ, ˆ w(µ)) ˆ − A(µ, w(µ)) w(µ) ˆ − w(µ)

  µˆ − µ Dµ A µ, ˆ w(µ) ˆ + o(1) + Lµˆ z(µ) + o(1) w(µ) ˆ − w(µ)

= Lµˆ z(µ) + o(1).

From this and a compactness argument and the uniqueness of w0 we easily deduce that z(µ) → w0 as µ → µ. ˆ Since w0 ≫P0 0, we find that for all µ < µˆ but close to µ, ˆ z(µ)
P0 (1/2)w0 and hence w(µ) ˆ − w(µ)
α(µ)w0 , µˆ − µ where α(µ) =

ˆ − w(µ) 1 w(µ) → +∞ 2 µˆ − µ

as µ → µˆ − 0.

This proves Step 3. S TEP 4. The set (µ∗ , µ0 ) \ O ∗ has measure 0 in R1 . Let l be a nontrivial positive functional on E0 , i.e., l ∈ P0∗ \ {0}, where P0∗ := {e ∈ E0∗ : e(w)
0 for w ∈ P0 }. Define f : (µ∗ , µ0 ) → R1 by f (µ) = l(w(µ)). Then f (µ) is increasing and hence has finite derivatives almost everywhere in (µ∗ , µ0 ). If µˆ ∈ (µ∗ , µ0 ) \ O ∗ , then by the conclusions in Steps 1–3, w(µ) is either discontinuous at µ, ˆ or it is continuous at µˆ but (3.16) holds. In either case we deduce (f (µ) ˆ − f (µ))/(µˆ − µ) → +∞ as µ → µˆ − 0. Therefore (µ∗ , µ0 ) \ O ∗ must have measure 0.  Let us now consider the stability of w(µ) := (uµ , v µ ) for µ ∈ [µ∗ , µ0 ) \ O ∗ . We have the following result. T HEOREM 3.13. Let µˆ ∈ (µ∗ , µ0 ) \ O ∗ . Then the following hold: (i) If w(µ) is continuous at µ, ˆ then w(µ) ˆ is asymptotically stable. (ii) If w(µ) is discontinuous at µ, ˆ then w(µ) ˆ is unstable. w(µ) is asymptotically (iii) If w(µ) is discontinuous at µ, ˆ then w(µˆ − ) := limµ→µ−0 ˆ stable. P ROOF. We first claim that r(Lµˆ ) = 1, where Lµˆ := Dw A(µ, ˆ w(µ)). ˆ Indeed, we can find ˆ It follows by the right-continuity of w(µ) on µ that r(Lµˆ ) = µn ∈ O ∗ decreasing to µ. / O ∗ , we must have r(Lµˆ ) = 1. This proves our claim. limn→∞ r(Lµn )  1. Since µˆ ∈

Bifurcation and related topics

161

We can now easily check that all the conditions in Theorem 1.3 are satisfied by ˆ w(µ)). ˆ For example, the condition F (µ, w) := w − A(µ, w) with (λ0 , v0 ) = (µ,



 ˆ w(µ) ˆ ∈ / R Fw µ, ˆ w(µ) ˆ Fµ µ,

follows from the last conclusion in Theorem 3.11, since Fµ (µ, ˆ w(µ)) ˆ = −(0, v µˆ ) ∈ µ ˆ ˆ w(µ)) ˆ = I − L . Hence the solutions of w − A(µ, w) = 0 (−P0 ) \ {(0, 0)} and Fw (µ, near (µ, ˆ w(µ)) ˆ form a smooth curve 

(µ, w) = µ(s), w(µ) ˆ + sw0 + τ (s) ,

s ∈ [−s0 , s0 ], s0 > 0,

with µ(0) = µ, ˆ µ′ (0) = 0, τ (0) = τ ′ (0) = 0, and w0 ≫P0 0 satisfies Lµˆ w0 = w0 . Since A is analytic, so are µ(s) and τ (s). We cannot have µ(s) ≡ 0 since by the right-continuity of w(µ),



 µ, w(µ) → µ, ˆ w(µ) ˆ

as µ → µˆ + 0.

(3.17)

Therefore there exists some integer k
2 such that µ′ (0) = · · · = µ(k−1) (0) = 0 and µ(k) (0) = 0. We claim that µ(k) (0) > 0, for if µ(k) (0) < 0 and k is even, then µ(s)  µˆ for all small |s|, and hence F (µ, w) = 0 has no solution near (µ, ˆ w(µ)) ˆ with µ > µ, ˆ contradicting (3.17); if µ(k) (0) < 0 and k is odd, then for s < 0 close to 0, we have µ(s) > µˆ and ˆ + sw0 + τ (s) ≪P0 w(µ), ˆ which implies that for µ > µˆ but close to µ, ˆ the only ws := w(µ) ˆ contradicting (3.17) solution (µ, w) of F (µ, w) = 0 close to w(µ) ˆ satisfies w ≪P0 w(µ), ˆ for µ > µ. ˆ Therefore we always have µ(k) (0) > 0. and the fact that w(µ) ≫P0 w(µ) If k is odd, then µ(s) crosses µˆ as s increases across 0. We show that w(µ) is continuous ˆ as at µ. ˆ Otherwise by Step 1 in the proof of Theorem 3.12, w(µ) → w(µˆ − ) ≪P0 w(µ) µ → µˆ − 0. This implies that w(µ(s)) ≪P0 ws for all s < 0 close to 0, contradicting the fact that w(µ(s)) is the maximal positive solution. This proves the continuity of w(µ) at µ. ˆ We may assume that s0 has been chosen small enough such that µ(s) < µˆ for s ∈ [−s0 , 0), µ(s) > µˆ for s ∈ (0, s0 ], and ˆ ws ≪P0 w(µ)

for s ∈ [−s0 , 0),

ws ≫P0 w(µ) ˆ for s ∈ (0, s0 ].

(3.18)

By the monotonicity of A, we have 

ˆ ws ) ws = A µ(s), ws P0 A(µ, ˆ ws )

for s ∈ [−s0 , 0), for s ∈ (0, s0 ].

ˆ Therefore {ws : s ∈ [−s0 , 0)} is a continuum of strict lower solutions of (3.1) with µ = µ, ˆ Now and {ws : s ∈ (0, s0 ]} is a continuum of strict upper solutions of (3.1) with µ = µ. the argument in the proof of Theorem 3.9 can be repeated, with ws , s ∈ [−s0 , s0 ] replacing (uδ , vδ ), δ ∈ [−δ0 , δ0 ], to conclude that w(µ) ˆ is asymptotically stable. If k is even, then µ(s)
µˆ for |s| small and hence there is no solution (µ, w) close to (µ, ˆ w(µ)) ˆ with µ < µ. ˆ Therefore w(µ) has to be discontinuous at µ. ˆ As before we

162

Y. Du

may assume that s0 is small enough such that (3.18) holds and µ(s) > µˆ for all s ∈ [−s0 , s0 ] \ {0}. Then

 ˆ ws ) ws = A µ(s), ws >P0 A(µ,

for s ∈ [−s0 , 0),

ˆ and hence {ws : s ∈ [−s0 , 0)} is a continuum of strict upper solutions of (3.1) with µ = µ. By an analogous consideration as in the proof of Theorem 3.9, we find that the unique solution of (3.13) with µ = µˆ and with initial data ws decreases in the order P0 as t increases, and it converges to a steady-state of (3.13) as t → ∞, say w∗ . Since w∗ P0 ws , ˆ is unstable. it follows by a sweeping argument that w∗ P0 w−s0 . This implies that w(µ) It remains to show conclusion (iii). So we assume that w(µ) is discontinuous at µ. ˆ ˆ w(µˆ − )). Since Let µn ∈ O ∗ be a sequence increasing to µˆ and denote L := Dw A(µ, w(µn ) → w(µˆ − ) and r(Lµn ) < 1, we deduce r(L) = limn→∞ r(Lµn )  1. If r(L) < 1, then the asymptotic stability of w(µˆ − ) follows from Theorem 3.10. Suppose next r(L) = 1. Then we can again apply Theorem 1.3 to conclude that the solutions of w = A(µ, w) near (µ, ˆ w(µˆ − )) form an analytic curve: 

(µ, w) = µ(s), w(µˆ − ) + sw0 + τ (s) ,

s ∈ [−s0 , s0 ], s0 > 0,

with µ(0) = µ, ˆ µ′ (0) = 0, τ (0) = τ ′ (0) = 0, and w0 ≫P0 0 satisfies Lw0 = w0 . Since w(µ) → w(µˆ − )

as µ → µˆ − 0,

(3.19)

we must have µ(s) ≡ 0, and hence there exists k
2 such that µ′ (0) = · · · = µ(k−1) (0) = 0 and µ(k) (0) = 0. We claim that k is odd and µ(k) (0) > 0. Indeed, if k is even and µ(k) (0) > 0, then µ(s)
µˆ for |s| small. This implies that w = A(µ, w) has no soluˆ contradicting (3.19). If k is either odd or even, tion close to (µ, ˆ w(µˆ − )) with µ < µ, (k) but µ (0) < 0, then for s > 0 small, µ(s) < µˆ and ws := w(µˆ − ) + sw0 + τ (s) ≫P0 w(µˆ − ) ≫P0 w(µ(s)), contradicting the fact that w(µ(s)) is the maximal solution of (3.1) with µ = µ(s). This proves our claim. Now the asymptotic stability of w(µˆ − ) can be proved as before, since (3.1) with µ = µˆ has a continuum of strict lower solutions  {ws : s ∈ [−s0 , 0)}, and a continuum of strict upper solutions {ws : s ∈ (0, s0 ]}. R EMARK 3.14. By standard regularity theory for parabolic equations, a steady-state (u0 , v0 ) of (3.13) is asymptotically stable in E0 implies that it is asymptotically stable in L∞ (Ω)2 . Under a slightly different definition for “asymptotically stable” solutions, the result of Theorem 3.13 was proved in [Da6] by a combination of fixed point index and local bifurcation argument. See also [Da7] for related results. We have parallel stability results for the minimal positive solutions (uµ , vµ ). More precisely, suppose that µ∗ > µ0 and for µ ∈ (µ0 , µ∗ ) denote Lµ := D(u,v) A(µ, uµ , vµ ),



 O∗ := µ ∈ µ0 , µ∗ : r(Lµ ) < 1 .

T HEOREM 3.15. The set (µ0 , µ∗ ) \ O∗ has measure zero in R1 .

Bifurcation and related topics

163

T HEOREM 3.16. z(µ) := (uµ , vµ ) is left continuous in (µ0 , µ∗ ). Moreover, for µˆ ∈ (µ0 , µ∗ ) \ O ∗ , the following hold. (i) If z(µ) is continuous at µ, ˆ then z(µ) ˆ is asymptotically stable. (ii) If z(µ) is discontinuous at µ, ˆ then z(µ) ˆ is unstable. z(µ) is an asymptotically (iii) If z(µ) is discontinuous at µ, ˆ then z(µˆ + ) := limµ→µ+0 ˆ stable positive solution of (3.1). Theorems 3.15 and 3.16 can be proved by analogous arguments to those used in the proofs of Theorems 3.12 and 3.13. We omit the details. R EMARK 3.17. If (uµ∗ , v µ∗ ) exists, i.e., (3.1) has a maximal positive solution with µ = µ∗ , then by the proof of Theorem 3.13, we easily see that it is unstable. Similarly, if (uµ∗ , vµ∗ ) exists, it is unstable. 3.3. Stable patterns We show that the global bifurcation branch of (3.1) discussed in Section 3.1 can be better described if ε is small. Moreover, we will show that, as ε → 0, µ∗ → ∞ and the minimal positive solution (uµ , vµ ) develops a sharp pattern determined by b(x). A key ingredient in our analysis here is the following degenerate model, that is, (3.1) with ε = 0  2   −
u = λu − b(x)u − cuv, 2 −
v = µv − v − duv,   u|∂Ω = v|∂Ω = 0.

(3.20)

We will regard (3.20) as a limiting problem for (3.1) with small ε. In a sense, our strategy here is similar to that of Section 2.2, where the perturbed problem (2.10) was studied by making use of the limiting problem (2.1); here we study the perturbed system (3.1) by its limiting problem (3.20). We will mainly follow [Du4]. We firstly apply a global bifurcation analysis to (3.20). The following a priori estimate is crucial to our analysis; we refer to [Du4], Theorem 2.1, for its proof, which is quite involved and uses some ideas in Section 2. T HEOREM 3.18. Given real numbers λ and M, there exists C = C(λ, M) > 0 such that any positive solution (u, v) of (3.20) with µ  M satisfies uL∞ (Ω) + vL∞ (Ω)  C. Ω

0 Let us observe that (3.20) behaves similarly to (3.1) if λΩ 1 < λ < λ1 . Indeed, we have a 0 trivial solution branch Γ0 := {(µ, 0, 0): µ ∈ R1 }, two semitrivial solution branches Γ10 := {(µ, uλ , 0): µ ∈ R1 } and Γ20 := {(µ, 0, θµ ): µ > λΩ 1 }, where uλ is the unique positive solution of (2.1) with p = 2. Moreover, we can apply the local and global bifurcation analysis of [BB2] as in Section 3.1 to conclude that there exists a global bifurcation branch

164

Y. Du

of positive solutions of (3.20), denoted by Γ 0 , which bifurcates from Γ10 at (µ˜ 0 , uλ , 0) 0 and joins Γ20 at (µ0 , 0, θµ0 ), where µ˜ 0 = λΩ 1 (duλ ) and µ is given by (3.9). We now find that all our discussions in the rest of Section 3.1 and in Section 3.2 can be applied Ω0 to (3.20) and yield the same results. Therefore we may conclude that when λΩ 1 < λ < λ1 , (3.20) and (3.1) behave similarly. It is also easily seen that they behave similarly when λ  λΩ 1 . We will show next that, when Ω

λ
λ1 0 ,

(3.21)

essential differences arise between (3.1) and (3.20). We henceforth assume that (3.21) holds. The first difference is that the semitrivial solution branch Γ10 disappears, though Γ00 and Γ20 are unchanged. We can still apply a local and global bifurcation analysis to (3.20): A local bifurcation analysis along Γ20 shows that a branch of positive solutions Γ 0 bifurcates from Γ20 at (µ0 , 0, θµ0 ) ∈ Γ20 ; by Theorem 3.18 and Rabinowitz’s global bifurcation theorem and the strong maximum principle, we find that Γ 0 is unbounded through µ becoming unbounded. Moreover, if (3.20) has a positive solution Ω 0 (u, v), then µ = λΩ 1 (v + du) > λ1 . Therefore Γ becomes unbounded through µ → +∞. 0 Γ can be further described by making use of the monotonicity property of (3.20) as in Sections 3.1 and 3.2, and we collect these results in the following theorem (see [Du3] and [Du4] for a detailed proof ). Ω

T HEOREM 3.19. Suppose λ
λ1 0 (0). Then: (i) (Existence and nonexistence.) There exists µ˜ ∗  µ0 such that (3.20) has no positive solution for µ < µ˜ ∗ , and it has at least one positive solution for µ > µ˜ ∗ . (ii) (Multiplicity and stability.) If µ˜ ∗ < µ0 , then (3.20) has at least two positive solutions for µ ∈ (µ˜ ∗ , µ0 ), and at least one positive solution for µ = µ˜ ∗ . Moreover, at least one positive solution is asymptotically stable for µ ∈ (µ˜ ∗ , µ0 ). (iii) (Continuum.) All the positive solutions of (3.20) stated in (i) and (ii) above can be chosen from the unbounded positive solution branch Γ 0 which joins the semitrivial solution (µ0 , 0, θµ0 ) and ∞. R EMARK 3.20. If d > 0 is small, then µ˜ ∗ < µ0 ; see [Du3], Theorem 3.6, for some estimates of µ˜ ∗ . We now come back to (3.1). To emphasize the dependence on ε, we denote the global positive solution branch of (3.1) by Γ ε , instead of Γ used in Section 3.1. Let us observe that the trivial solution branch Γ0 and the semitrivial solution branch Γ2 = {(µ, 0, θµ )} are independent of ε, as is µ0 given by (3.9); but Γ1 is ε-dependent and we henceforth denote it by Γ1ε = {(µ, φλε , 0)}. Similarly, we replace µ∗ and µ∗ by µ∗ (ε) and µ∗ (ε), respectively. ε We also replace µ0 by µ0 (ε), which, we recall, is given by µ0 (ε) := λΩ 1 (dφλ ). As ε → 0, ε the behavior of φλ is described by Theorem 2.6. We now consider the behavior of µ0 (ε) (see [Du3] and [Du4] for a proof ).

Bifurcation and related topics

165 Ω

+ ε P ROPOSITION 3.21. As ε → 0, µ0 (ε) = λΩ 1 (dφλ ) → λ1 (dU λ ), which is finite. Here 0 , U λ is the minimal positive solution of (2.3) with p = 2, and λΩ+ (dU λ ) is Ω+ := Ω \ Ω 1 defined by

 Ω Ω λ1 + (dU λ ) := lim λ1 + min{n, dU λ } . n→∞

Next we study the changes in µ∗ (ε) and µ∗ (ε) as ε → 0. P ROPOSITION 3.22. The functions ε → µ∗ (ε) and ε → µ∗ (ε) are both nonincreasing. Moreover, limε→0 µ∗ (ε) = ∞ and limε→0 µ∗ (ε) = µˆ ∗  µ˜ ∗ , where µ˜ ∗ is defined in Theorem 3.19. P ROOF. We first show that ε → µ∗ (ε) is nonincreasing. If µ∗ (ε) ≡ max{µ0 , µ0 (ε)}, then, ε ∗ since µ0 (ε) = λΩ 1 (dφλ ) is nonincreasing with ε, there is nothing to prove. If µ (ε) > max{µ0 , µ0 (ε)} for some ε = ε0 > 0, then by Theorem 3.6, (3.1) with ε = ε0 and µ = µ∗ (ε0 ) has a positive solution (u0 , v0 ). Let ε1 ∈ (0, ε0 ]. Then

 −
u0  λu0 − b(x) + ε1 u20 − cu0 v0 .

Hence (u0 , v0 ) is an upper solution to (3.1) with ε = ε1 . Since µ = µ∗ (ε0 ) > µ0 (ε0 ), if we choose ε1 close enough to ε0 , then µ > µ0 (ε1 ) and hence the problem −
v = µv − v 2 − dφλε1 v,

v|∂Ω = 0,

has a unique positive solution vε1 . It is easily checked that (φλε1 , vε1 ) is a lower solution ε to (3.1) with ε = ε1 . Moreover, it is easily seen that φλε1
φλ0
u0 and vε1  v0 . Thus, by standard upper and lower solution argument for competition models, (3.1) with ε = ε1 has a positive solution (u, v) satisfying u0  u  uε1 and vε1  v  v0 . By the definition of µ∗ (ε), we must have µ∗ (ε1 )
µ = µ∗ (ε0 ). Thus ε → µ∗ (ε) is always nonincreasing. The fact that ε → µ∗ (ε) is nonincreasing can be proved by a similar argument. Next we prove that µ∗ (ε) → ∞ as ε → 0. We view (3.1) as a perturbation of (3.20) and Ω use a degree argument. Given any µ˜ > max{λ1 + (dU ), µ˜ ∗ }, by Theorem 3.18, we can find a constant C > 0 such that any positive solution (u, v) of (3.20) with µ ∈ [0, µ] ˜ satisfies u∞  C. Note also that we always have v  θµ  θµ˜ . We now recall the definition of A(µ, u, v) in Section 3.1. By enlarging ξ, η and M there, and replacing µˆ there by µ˜ if µ˜ > µ, ˆ we find that the properties of A are retained for all small ε
0. To emphasize the dependence on ε, we denote A(µ, u, v) by Aεµ (u, v).  × To use a fixed point index argument, it is convenient to work in the space E := C(Ω)  C(Ω) with the natural positive cone P := {(u, v) ∈ E: u
0, v
0}. If we define A := {(u, v) ∈ E: 0  u < ξ, 0  v < η}, then it is easily checked that Aεµ : A → P and is completely continuous for all small ε
0 and µ ∈ [0, µ]. ˜ Let us now consider the fixed point index, indexP (A0µ , A). When λΩ < µ < µ ˜ , the only nonnegative solutions of (3.20) ∗ 1 are (u, v) = (0, 0) and (u, v) = (0, θµ ), both are linearly unstable solutions of (3.20). By Dancer’s fixed point index formula [Da2], for such µ,



 indexP A0µ , (0, 0) = indexP A0µ , (0, θµ ) = 0.

166

Y. Du

Therefore, 





indexP A0µ , A = indexP A0µ , (0, 0) + indexP A0µ , (0, θµ ) = 0.

As A0µ has no fixed point on ∂P A (the relative boundary of A with respect to P ) for any µ ∈ [0, µ], ˜ by the continuity property of the fixed point index (see [Am]), indexP (A0µ , A) is independent of µ ∈ [0, µ] ˜ and is thus identically zero. Consider now µ ∈ (µ0 , µ]. ˜ For such µ, the trivial solution (0, 0) of (3.20) is linearly unstable and hence has fixed point index 0, but the semitrivial solution (0, θµ ) is linearly stable, and therefore it has fixed point index 1. It follows that we can find small neighborhoods N0 of (0, 0) and N1 of (0, θµ˜ ) such that

 indexP A0µ˜ , A \ (N0 ∪ N1 )





 = indexP A0µ˜ , A − indexP A0µ˜ , (0, 0) − indexP A0µ˜ , (0, θµ˜ ) = 0 − 0 − 1 = −1.

It follows from the continuity property of the fixed point index that, for all sufficiently small ε > 0, indexP (Aεµ˜ , A \ (N0 ∪ N1 )) is well defined and equals indexP (A0µ˜ , A \ (N0 ∪ N1 )) = −1. Thus Aεµ˜ has a fixed point (u, v) in A \ (N0 ∪ N1 ), i.e., (3.20) has a positive ˜ As µ˜ is arbitrary, this solution with µ = µ˜ for all small ε > 0. In particular, µ∗ (ε)
µ. implies µ∗ (ε) → ∞ as ε → 0. Let us now prove that limε→0 µ∗ (ε)  µ˜ ∗ . Since µ∗ (ε) is nonincreasing with ε and µ∗ (ε)  µ0 , µˆ ∗ := limε→0 µ∗ (ε) exists. If µˆ ∗ > µ˜ ∗ , then since µˆ ∗  µ0 , we must have µ∗ < µ0 . By Theorem 3.19, (3.20) with µ = µ˜ ∗ has a positive solution (u0 , v0 ). It is easily checked that (u0 , v0 ) is a lower solution to (3.1) with µ = µ˜ ∗ for any ε > 0. Moreover, Ω since µ˜ ∗ < µ0 , λ = λΩ 1 (cθµ0 ) > λ1 (cθµ˜ ∗ ), and thus, the problem

 −
u = λu − b(x) + ε u2 − cθµ˜ ∗ u,

u|∂Ω = 0,

has a unique positive solution u∗ . Clearly (u∗ , θµ˜ ∗ ) is an upper solution of (3.1) with µ = µ˜ ∗ , and v0  θµ˜ ∗ , u0
u∗ . Thus, (3.1) with µ = µ˜ ∗ has a positive solution, and hence µ∗ (ε)  µ˜ ∗ . But this implies µˆ ∗  µ˜ ∗ , a contradiction. Therefore, we must have µˆ ∗  µ˜ ∗ , as required.  By Propositions 3.21 and 3.22, we find that, for small ε > 0, µ0 (ε) < µ∗ (ε), and for Ω fixed µ > λ1 + (dU λ ), µ ∈ (µ0 (ε), µ∗ (ε)) holds for all small ε > 0. Therefore, (3.1) has a minimal positive solution (uεµ , vµε ). We are interested in the profile of (uεµ , vµε ) as ε → 0. To simplify notation, we will denote this minimal solution by (uε , v ε ) when its dependence on µ is not emphasized. For technical reasons, in the following theorems we require µ > Ω λ ), where U λ is the maximal positive solution of (2.3) with p = 2. Under some λ1 + (d U λ = U λ , see Remark 2.3. mild conditions on b(x), U Ω Ω λ ). Then the following concluT HEOREM 3.23. Suppose that λ > λ1 0 and µ > λ1 + (d U sions hold:

Bifurcation and related topics

167

0 , (i) (uε , v ε ) → (∞, 0) uniformly on Ω µ  lim ε→0 uε , limε→0 uε  U µ , V µ  lim ε→0 v ε , limε→0 v ε  V µ , where the (ii) U  µ ) are re  limits are uniform on any compact subset of Ω \ Ω0 , ( U µ , V µ ) and (Uµ , V spectively the minimal and maximal positive solutions of the boundary blow-up problem  2 x ∈ Ω+ ,   −
u = λu − b(x)u − cuv, 2 (3.22) −
v = µv − v − duv, x ∈ Ω+ ,   u|∂Ω = v|∂Ω = 0, u|∂Ω0 = ∞, v|∂Ω0 = 0.

Moreover, for any positive sequence {εn } that converges to 0, {(uεn , v εn )} has a subse\Ω 0 , to a positive solution quence that converges, uniformly on any compact subset of Ω of (3.22).

We omit the rather involved proof of Theorem 3.23 here, and refer the interested reader to [Du4]. Let us note that for small ε > 0, the above result shows that (uε , v ε ) exhibits a sharp pattern over the underlying domain Ω: uε is large over Ω0 and is positive and of order 1 over Ω+ ; v ε is small over Ω0 , and it is positive and of order 1 over Ω+ . Our next result demonstrates that an intuitively clearer pattern is given by a rescaled version of (uε , v ε ), namely (u˜ ε , v ε ) := (εuε , v ε ). It is easily checked that (u˜ ε , v ε ) is a minimal positive solution of the following competition model  % −1 & 2   −
u = λu − ε b(x) + 1 u − cuv, (3.23) −
v = µv − v 2 − ε −1 duv,   u|∂Ω = v|∂Ω = 0. Clearly (u˜ ε , v ε ) has the same stability properties as (uε , v ε ) in (3.1) when regarded as a steady-state of the corresponding parabolic problem of (3.23). Ω

Ω

λ ). Then the following are true: T HEOREM 3.24. Suppose that λ > λ1 0 and µ > λ1 + (d U ε ε ˜  ˜ (i) (u˜ , v ) → (θλ , 0) uniformly on Ω0 , where, θλ is the unique positive solution of −
u = λu − u2

in Ω0 ,

u|∂Ω0 = 0.

(ii) For any positive sequence εn → 0, {(u˜ εn , v εn )} has a subsequence that converges to \Ω 0 , where V ∈ C(Ω \Ω 0 ) and is the second component of some (0, V ) uniformly on Ω positive solution (U, V ) of (3.22). Again we refer the proof of Theorem 3.24 to [Du4]. The above two theorems give us a rather detailed description of the spatial pattern of the minimal positive solution (uεµ , vµε ) with small ε > 0. By Theorem 3.16, this solution is asymptotically stable exactly when its dependence on µ is continuous at µ, which is the case for almost all µ. If its dependence on µ is discontinuous at some µ, ˆ then

ε ε

ε ε (3.24) uˆ , vˆ := lim uµ , vµ µ→µ+0 ˆ

168

Y. Du

Fig. 1. Bifurcation diagram for (3.1) and (3.20).

is a positive solution of (3.1) with µ = µ, ˆ and it is asymptotically stable. Equation (3.24) implies that, for fixed small ε > 0, (uˆ ε , vˆ ε ) has the spatial pattern similar to (uεµ , vµε ) with µ > µˆ but close to µ. ˆ The bifurcation diagram (Figure 1) describes a possible scenario of the global bifurcation Ω branches for (3.1) and (3.20) with λ
λ1 0 .

3.4. Remarks 1. If Ω0 is not connected but consists of finitely many components, then our results in Sections 3.1 and 3.2 are not affected, and the results in Section 3.3 can also be extended to this case, see [Du4] for details. 2. The strategy employed here for the competition model (3.1) has been used to study certain predator–prey models; see [DD2] and [DHs] for details, and [Du5] for a survey. Since the predator–prey models do not have any kind of monotonicity property, the techniques there are very different from here. 3. It is an interesting problem to see what new features arise if we have at least two nonconstant coefficients in (3.1) that are close to zero in certain parts of the domain. 4. Our method works as well if (3.1) has the following more general form   % & 2

  − div d1 (x)u = λa1 (x)u − b(x) + ε u − c(x)uv,

 − div d2 (x)v = µa2 (x)v − e(x)v 2 − d(x)uv,   Bu|∂Ω = Bv|∂Ω = 0,

(3.25)

Bifurcation and related topics

169

where the coefficient functions are positive except b(x) which is allowed to vanish on part of the domain, and the boundary operator B is either Dirichlet, or Neumann or Robin type. 5. In a series of recent papers, Hutson–Lou–Mischaikow–Polacik studied various perturbations of the special competition model  2   ut − µ
u = α(x)u − u − uv, (x, t) ∈ Ω × (0, ∞), vt − µ
v = α(x)v − v 2 − uv, (x, t) ∈ Ω × (0, ∞),   uν = vν = 0, (x, t) ∈ ∂Ω × (0, ∞),

and obtained interesting results revealing some fundamental effects of heterogeneous environment on the competition model. We refer to [HLM,HLMP] and [HMP] for details. Further related results can be found in [AC,CC,CCH,Lop1,LS]. 6. Several general approaches have been developed in the past two decades to study (3.25). For example, the method of monotone iterations and order preserving operators was developed and used by Pao [Pao], Koman and Leung [KL] and many others; the method of global bifurcation was introduced by Blat and Brown [BB1,BB2]; and the fixed point index approach was developed by Dancer [Da3,Da4]. These methods were applied to (3.25) with constant coefficients, but they work as well with nonconstant coefficients to yield similar results. Therefore it is difficult to rely on these methods alone to reveal the effects of heterogeneous environment on (3.25). 7. The spatial behavior of positive solutions of the two species competition model has received extensive studies even in the constant coefficient case, i.e., when the spatial environment is homogeneous. In [KW], it was shown that, if the spatial domain Ω is convex, then problem (3.25) with constant coefficients and Neumann boundary conditions has no stable positive steady-state that depends on x, i.e., all its stable positive steady-states are constant solutions. On the other hand, in [MM], spatially variable stable positive steadystate solutions were constructed for certain nonconvex Ω (see also [KY]). In [DD1], it was proved that in the strong competition case, positive steady-states of (3.25) with constant coefficients tend to segregate over Ω, i.e., uv is close to 0 with u close to max{w, 0} and v close to max{−w, 0}, where w is a sign-changing solution of a scalar elliptic equation deduced from this competition system. In [LN2,LN1], the competition model with self-diffusion and cross-diffusion was closely examined and the existence and asymptotic profile of space dependent positive steady-states were obtained when certain parameters are large; see also [Mi] and [MK] for earlier result. 8. It is more realistic to assume that the coefficients in the competition model are also dependent on time, for example, they are periodic in time as well as a function of the space variable x. The general case was systematically discussed in [He]. It would be interesting to see whether the results of Sections 3.1–3.3 here can be extended to this case.

4. Bifurcation and exact multiplicity: The perturbed Gelfand equation Exact multiplicity of solutions to nonlinear equations is in general very difficult to obtain. In most nonlinear elliptic problems, the number of solutions is not only affected by the nonlinearity in the equation, it also depends on the geometry of the underlying domain

170

Y. Du

(see [Da5]). When the underlying domain is a ball, then for many classes of nonlinearities, it is possible to use a bifurcation approach to find the exact number of positive solutions (see, e.g., [OS1,OS2]). A key point is that, by the well-known result of Gidas, Ni and Nirenberg [GNN], under homogeneous Dirichlet boundary conditions, any positive solution on the ball is radially symmetric; this reduces the PDE problem to an ODE one and makes the exact multiplicity problem reachable. We will demonstrate this approach through the perturbed Gelfand equation, where a perturbation argument is also needed. We mainly follow [DLo2] and [Du2]. The perturbed Gelfand equation arises in the mathematical modelling of thermal reaction processes (see [BE], Section 1.3), which is of the following form: −
u = λeu/(1+εu)

in Ω,

u|∂Ω = 0,

(4.1)

where Ω is a bounded smooth domain in RN , λ is a positive constant known as the Frank– Kamenetskii parameter, ε > 0 is a small parameter representing the reciprocal activation energy and u stands for the dimensionless temperature. If ε = 0, problem (4.1) reduces to the well-known Gelfand equation −
u = λeu

in Ω,

u|∂Ω = 0.

(4.2)

If we let v = ε 2 u and µ = λε 2 e1/ε , then (4.1) becomes −
v = µe−1/(ε+v)

in Ω,

v|∂Ω = 0,

(4.3)

which has the limiting equation −
v = µe−1/v

in Ω,

v|∂Ω = 0.

(4.4)

We will make use of both (4.2) and (4.3) to obtain a good understanding of (4.1) for small ε > 0. If ε
1/4, then it is easily checked that the right-hand side of (4.1) is an increasing concave function of u, and it follows easily that (4.1) has a unique positive solution for every λ > 0; see [BIS] and [CS]. For a general ε > 0, it is known that (4.1) has a unique positive solution for 0 < λ ≪ 1 and λ ≫ 1; and for 0 < ε ≪ 1, it is known that there exists a nonempty bounded open interval Λ ⊂ (0, ∞) such that for λ ∈ Λ, (4.1) has at least three distinct positive solutions u1 (x)  u2 (x)  u3 (x); see [BIS,CS,Sh,Wie1] and [Wie2]. These results suggest that when ε > 0 is small, the global bifurcation branch {(λ, u)} of (4.1) is roughly S-shaped. If the space dimension is one, then it is proved by the method of quadratures that this bifurcation branch is a continuous curve and is exactly S-shaped when ε > 0 is sufficiently small [HM], and it was further shown that this is true when 0 < ε < 1/4.4967 in [W1], and when 0 < ε < 1/4.35 in [KLi]. It had been conjectured that the global bifurcation curve of (4.1) is exactly S-shaped when the space dimension is two and Ω is a ball. (This kind of result is useful in understanding the profiles of the solutions to the full exothermic reaction–diffusion system; see [MS] for details.) Parter [Pa] considered this case for the equivalent problem (4.3), and

Bifurcation and related topics

171

gave estimates of four positive values µ 1 (ε) < µ¯ 1 (ε) < µ 2 (ε) < µ¯ 2 (ε) such that (4.3) has a unique positive solution if µ ∈ (0, µ 1 (ε)] ∪ [µ¯ 2 (ε), +∞) and it has at least three positive solutions if µ ∈ [µ¯ 1 (ε), µ 2 (ε)]. By using (4.2) as a limiting problem for (4.1), Dancer [Da1] proved, among other things, that for any small positive λ0 > 0, one can find an ε0 > 0 small such that if ε ∈ (0, ε0 ), then there is a constant λ2 (ε) > 0 such that (4.1) has exactly three positive solutions if λ ∈ (λ0 , λ2 (ε)); it has exactly two positive solutions if λ = λ2 (ε); and there is a unique positive solution if λ > λ2 (ε). This leaves the conjecture unsolved only for the small λ-range, 0 < λ < λ0 . This gap was finally filled in [DLo2], where apart from (4.2), the limiting equation (4.4) was also used; we will give a proof of this result further based on (4.4) only. As will become clear later, when the space dimension N is greater than two, the global solution curve of (4.1) needs not be S-shaped for small ε > 0; it is more complicated when 3  N  9, and (4.2) will play an important role in understanding this.

4.1. The limiting equations When Ω is the unit ball, the number of positive solutions to (4.2) was completely described in the well-known paper of Joseph and Lundgren [JL] based on a phase plane method. Their results are summarized in the following proposition. P ROPOSITION 4.1. Suppose that Ω is the unit ball in RN . Then there exists a finite value λ∗ depending on N , such that (4.2) has (i) no positive solution when λ > λ∗ (1  N  9), (ii) exactly one positive solution when λ = λ∗ (1  N  9), (iii) exactly two positive solutions when 0 < λ < λ∗ (N = 1, 2), (iv) an infinite number of positive solutions when λ = 2(N − 2) (3  N  9), (v) a finite but large number of solutions when |λ − 2(N − 2)| = 0 is small (3  N  9), (vi) a unique positive solution when 0 < λ < 2(N − 2), and no positive solution for λ
2(N − 2) (N
10). Moreover, the solution set {(λ, u)} is a smooth curve which can be illustrated in Figure 2.

Fig. 2. Bifurcation diagrams for the Gelfand equation (4.2) over the unit ball.

172

Y. Du

We now consider the other limiting equation (4.4). We will use a bifurcation argument to show that if Ω is a ball of any dimension, the positive solution curve {(µ, v)} is exactly “⊂”-shaped. This is in sharp contrast to the Gelfand equation (4.2), whose positive solution set changes structure as the dimension of the underlying domain changes. For definiteness, we assume that Ω is the unit ball B := {x ∈ RN : |x| < 1}. For convenience of notation, we also denote f (v) = f0 (v) = e−1/v ,

fε (v) = e1/(ε+v) .

(4.5)

The following lemma, with a difficult and technical proof, is crucial. L EMMA 4.2. Suppose Ω = B. If u is a degenerate positive solution of (4.4), that is, the linearized problem −
φ = µf ′ (u)φ,

φ|∂B = 0,

has a nontrivial solution φ, then φ does not change sign in B. P ROOF. By [GNN], u is radially symmetric: u(x) = u(r), r = |x|; moreover, u′ (r) < 0 on (0,1]. By Proposition 3.3 of [LN], φ is also radially symmetric, φ(x) = φ(r). Hence φ ′′ +

N −1 ′ φ + µf ′ (u)φ = 0 r

in [0, 1],

φ ′ (0) = 0, φ(1) = 0.

By the Harnack inequality (or a well-known uniqueness result for the above singular second-order ordinary differential equation), φ(0) = 0. We may assume φ(0) > 0. Direct calculations give f ′ (u) = u−2 e−1/u > 0 ∀u > 0,

f ′′ (u) = u−4 e−1/u (1 − 2u). Hence

f ′′ (u) > 0 for u ∈ (0, 1/2),

f ′′ (u) < 0 for u ∈ (1/2, ∞).

One easily sees K(u) =

uf ′ (u) 1 = f (u) u

is a decreasing function of u on (0, ∞). We divide our following discussion into two cases: (i) u(0)  1 and (ii) u(0) > 1.

Bifurcation and related topics

Consider case (i) first.

173

Let

v(r) = rur (r) + ξ u(r), where ξ is a positive constant to be specified later. Then

Define

%

& −
v − µf ′ (u)v = µ 2f (u) − ξ f ′ (u)u − f (u) %

& = µf (u) 2 − ξ K(u) − 1 . h(r) = −

ru′ (r) , u(r)

(4.6)

r ∈ [0, 1).

Clearly h(0) = 0 and h(1) = +∞. We show in the following that h′ (r) > 0 in (0, 1). Indeed, h′ (r) =

ru2r + (N − 2)ur u + µrf (u)u 2H (r) + µrQ(u(r)) = , u2 u2

(4.7)

where

 ru2r + (N − 2)ur u + µrF u(r) , 2  u F (u) = f (s) ds, Q(u) = uf (u) − 2F (u). H (r) =

0

Here and in what follows, ur is sometimes used for u′ to avoid notation like u′ 2 . If N = 1, 2, then it follows from the first equality in (4.7) that h′ (r) > 0 in (0, 1). Therefore, we need only consider N > 2. A simple calculation gives % N −1 &′

 r H (r) = µr N −1 G u(r)

with G(u) = N F (u) −

Clearly, G(0) = 0 and

N +2 N −2 ′ f (u) − f (u)u 2 2 (N + 2)u − (N − 2) = u−1 e−1/u . 2

G′ (u) =

It follows that G′ (u) < 0 on [0, γ ),

G′ (u) > 0 on (γ , ∞),

N −2 f (u)u. 2

174

Y. Du

where γ = (N − 2)/(N + 2). Therefore, we have either G(u) < 0 on (0, ∞) or G(u) < 0 on (0, γ0 ) and G(u) > 0 on (γ0 , ∞) for some γ0 > γ . We show that actually only the latter alternative can occur. Indeed, if G(u(r))  0 for all r ∈ [0, 1], then 0
γ0

(4.8)

whenever N > 2 and u is a positive solution of (4.4) with Ω = B. Let t be uniquely determined by u(t) = γ0 . We have % N −1 &′

 r H (r) = µr N −1 G u(r) > 0 ∀r ∈ (0, t),

which implies H (r) > 0 for r ∈ (0, t]. Moreover, for r ∈ (t, 1], G(u(r)) < 0 and therefore r N −1 H (r) = H (1) −
H (1) =



1 r

 µr N −1 G u(r) dr

u2r (1) > 0. 2

Thus we always have H (r) > 0 on (0, 1]. Since Q(0) = 0 and

% & Q′ (u) = uf ′ (u) − f (u) = f (u) K(u) − 1
0 ∀u ∈ [0, 1),

we have Q(u)
0 on [0, 1) and hence, by (4.7), h′ (r) > 0 on (0, 1), as required. Denote µ(r) =

2 . K(u(r)) − 1

Then µ(r) is strictly decreasing for r ∈ (0, 1], and by (4.6), −
v − µf ′ (u)v = g(r),

% &% & g(r) = µf (u) K(u) − 1 µ(r) − ξ .

With these preparations, we are now ready to show that φ does not change sign in B in case (i). We argue indirectly. Suppose φ(r) has a zero in (0, 1). Then we can find 0 < t1  t2 < 1 such that φ(t1 ) = 0,

φ(r) > 0 ∀r ∈ [0, t1 ),

φ(t2 ) = 0,

φ(r) = 0 ∀r ∈ (t2 , 1).

Bifurcation and related topics

175

Now we choose ξ = h(t1 ) in v = rur + ξ u, and have two subcases to consider (a) µ(t1 )
ξ

and (b) µ(t1 ) < ξ.

We have % & v(r) = rur + h(t1 )u = u h(t1 ) − h(r) > v(t1 ) = 0 ∀r ∈ [0, t1 ), v(r) < 0 ∀r ∈ (t1 , 1).

(4.9)

In subcase (a), we easily see g(r) > 0 on (0, t1 ), and hence, using v(t1 ) = 0, we arrive at the following contradiction 0
0 on (t2 , 1) for otherwise we can replace φ by −φ. Moreover, one easily sees g(r) < 0 on [t1 , 1]. Then by (4.9) and φ ′ (t2 ) > 0 > φ ′ (1), we also arrive at a contradiction 0>



1

g(r)φ(r)r N −1 dr

t2

= =



B\Bt2



∂Bt2

% & −
v − µf ′ (u)v φ

−vφr +



vφr > 0.

(4.11)

∂B

This proves the lemma for case (i). Next we consider case (ii) where u(0) > 1. We can find 0 < r1 < r2 < 1 uniquely determined by u(r1 ) = 1,

u(r2 ) = 1/4.

We first show φ(r) = 0 on (0, r1 ]. To this end, we choose w(r) = u(r) − 1/4 as a test function. Clearly −
w − µf ′ (u)w = µq(u), We have q ′ (u) = (1/4 − u)f ′′ (u),

q(u) = f (u) − f ′ (u)(u − 1/4).

176

Y. Du

which is positive on (0, 1/4), negative on (1/4, 1/2) and positive on (1/2, ∞). Since q(0) = 1/4f ′ (0) = 0, it follows q(u)
q(1/2) = f (1/2) − 1/4f ′ (1/2) = 0 ∀u > 0. Hence −
w − µf ′ (u)w = µq(u)
0 on B. If φ(r) has a zero in (0, r1 ], then we can find t ∈ (0, r1 ] such that φ(r) > 0 on [0, t) and φ(t) = 0. Using w(t) = u(t) − 1/4 > u(r2 ) − 1/4 = 0, we deduce 0



0

t

 µq u(r) φ(r)r N −1 dr =



Bt

% & −
w − µf ′ (u)w φ =



wφr < 0.

∂Bt

This contradiction finishes our proof for φ(r) = 0 on [0, r1 ]. Next we suppose φ(r) changes sign in (0, 1) and deduce a contradiction. Since φ(r) = 0 in [0, r1 ], we can find r1 < t1  t2 < 1 such that φ(t1 ) = 0,

φ(r) > 0 ∀r ∈ [0, t1 ),

φ(t2 ) = 0,

φ(r) = 0 ∀r ∈ (t2 , 1).

We now define v(r), h(r), µ(r) and g(r) as in case (i), and choose ξ = h(t1 ) in the definition of v(r). Since u(r)  1 on [r1 , 1], our arguments in case (i) give h′ (r) > 0 on (r1 , 1). Hence % & v(r) = u h(t1 ) − h(r) > v(t1 ) = 0 ∀r ∈ [r1 , t1 ),

v(r) < 0 ∀r ∈ (t1 , 1].

If µ(t1 )
ξ , then since µ(r) is strictly decreasing on (0, 1) and K(u(r)) < 1 on [0, r1 ), K(u(r)) > 1 on (r1 , 1], we have

and

% &% & g(r) = µf (u) K(u) − 1 µ(r) − ξ > 0 ∀r ∈ (r1 , t1 ) %

& g(r) = µf (u) 2 − ξ K(u) − 1 > 0

∀r ∈ [0, r1 ].

Thus g(r) > 0 on [0, t1 ). Now we can deduce the same contradiction (4.10) as in case (i). If µ(t1 ) < ξ , then % &% & g(r) = µf (u) K(u) − 1 µ(r) − ξ < 0 ∀r ∈ (t1 , 1],

Bifurcation and related topics

177

and we arrive at the contradiction (4.11) as in case (i). This finishes the proof of Lemma 4.2.  Using Lemma 4.2 and the turning point theorem of Crandall and Rabinowitz [CR1,CR2], Theorem 1.3, we can prove the following result. L EMMA 4.3. Suppose that u0 is a degenerate positive solution of (4.4) with Ω = B and  µ = µ0 . Then all the positive solutions (µ, u) of (4.4) that are near (µ0 , u0 ) in R1 × C(B) lie on a smooth curve represented by 

(µ, u) = µ0 + τ (s), u0 + sφ0 + z(s) with |s| small, where z(0) = z′ (0) = 0, τ (0) = τ ′ (0) = 0 and φ0 is a positive eigenfunction given in Lemma 4.2. Moreover, τ ′′ (0) > 0.

 Y = C α (B)  and F (µ, u) =
u + µf (u). It is easy to see that P ROOF. Set X = C02,α (B), F is a smooth mapping from R1 × X to Y . The partial derivative Fu at (µ0 , u0 ) is given by Fu (µ0 , u0 )φ =
φ + µ0 f ′ (u0 )φ. By Lemma 4.2 we know that N (Fu (µ0 , u0 )) is of one dimension: in fact, N (Fu (µ0 , u0 )) = span{φ0 }. Moreover, codim Fu (µ0 , u0 ) = 1 by the Fredholm alternative. Also

 / R Fu (µ0 , u0 ) Fµ (µ0 , u0 ) = f (u0 ) ∈  since B f (u0 )φ0 > 0. Therefore we can use Theorem 1.3 to conclude the following:

Near the degenerate solution (µ0 , u0 ) in R1 × X, the solutions of (4.4) form a smooth curve

  µ(s), u(s) = µ0 + τ (s), u0 + sφ0 + z(s) , (4.12)

where s → (τ (s), z(s)) ∈ R1 × Z is a smooth function near s = 0 with τ (0) = τ ′ (0) = 0, z(0) = z′ (0) = 0, where Z is a complement of span{φ0 } in X.

By a standard elliptic regularity consideration, we know that any solution of (4.4) near  is also close to (µ0 , u0 ) in R1 × X. Therefore all the positive (µ0 , u0 ) in R1 × C(B)  are contained in the smooth curve (4.12). solutions of (4.4) near (µ0 , u0 ) in R1 × C(B) We now substitute the expression (4.12) for the solutions into equation (4.4), differentiate the equation with respect to s twice at s = 0, multiply the resulting identity by φ and integrate it over B to obtain 1

f ′′ (u0 )φ03 r N −1 dr τ (0) = −µ0 0 1 . N −1 dr 0 f (u0 )φ0 r ′′

It remains to show that τ ′′ (0) > 0. Let use define % & ξ(r) = µ0 r N −1 f (u0 )φ0′ − f ′ (u0 )φ0 u′0 .

(4.13)

178

Y. Du

One easily checks that

2 ξ ′ (r) = −µ0 r N −1 f ′′ (u0 )φ0 u′0 ,

ξ(0) = ξ(1) = 0.

(4.14)

It follows that  1

2 f ′′ (u0 )φ0 u′0 r N −1 dr = 0. 0

We claim that f ′′ (u0 (r)) changes sign exactly once in (0, 1). Otherwise we necessarily have u0 (0)  1/2 and f ′′ (u0 (r)) > 0 in (0, 1). By the first part of (4.14), this implies that ξ ′ (r) < 0 in (0, 1), and hence ξ(0) > ξ(1), contradicting the second part of (4.14). This proves our claim. Let us assume that f ′′ (u0 (r)) changes sign exactly once in (0, 1) at r = r¯ . Next we show that −u′0 and φ0 intersect exactly once in (0, 1). Since −u′0 (0) = 0, −u′0 (1) > 0, φ0 (0) > 0 and φ0 (1) = 0, −u′0 and φ0 intersect at least once in (0, 1). If they intersect at least twice, we may assume that there exist 0 < r1 < r2 < 1 such that (φ0 + u′0 )(r1 ) = (φ0 + u′0 )(r2 ) = 0, and (φ0 + u′0 )(r) < 0 for r ∈ (r1 , r2 ). This in particular implies that (φ0 + u′0 )′ (r1 )  0 and (φ0 + u′0 )′ (r2 )
0. By the equations of φ0 and u′0 , it is easy to check that the following identity holds: &′ % N −1 ′ ′ r u0 φ0 − r N −1 φ0 u′′0 = −(N − 1)r N −2 u′0 φ0 .

(4.15)

Integrating (4.15) from r1 to r2 , and using (φ0 + u′0 )(r1 ) = (φ0 + u′0 )(r2 ) = 0, we obtain ′ ′

r2N −1 u′0 (r2 ) φ0 + u′0 (r2 ) − r1N −1 u′0 (r1 ) φ0 + u′0 (r1 )  r2 =− (N − 1)r N −2 u′0 φ0 dr.

(4.16)

r1

However, the right-hand side of (4.16) is positive since u′0 < 0 and φ0 > 0, while the lefthand side of (4.16) is nonpositive due to the facts summarized ahead of (4.15). This contradiction shows that −u′0 and φ0 intersect exactly once in (0, 1). By replacing φ0 by η0 φ0 for some suitable positive η0 if necessary, we may assume that −u′0 and φ0 intersect exactly at r = r¯ , where f ′′ (u0 (r)) changes sign. It now follows that f ′′ (u0 )φ02 < f ′′ (u0 )u′02

in (0, r¯ ) ∪ (¯r , 1).

Hence 

0

Since

1 0

1

f ′′ (u0 )φ03 r N −1 dr
0, the fact that τ ′′ (0) > 0 now follows from (4.13).

Before proving our main result for (4.4), we need one more preparation.



Bifurcation and related topics

179

L EMMA 4.4. Suppose g ∈ C 1 (R1 ) and B is the unit ball in RN , N
1. Then for any given c > 0, the problem −
u = λg(u),

u|∂B = 0,

can have at most one solution (λ, u) satisfying λ > 0, u
0 and u(0) = c. P ROOF. Suppose that (λ0 , u0 ) is such a solution. It suffices to show that any other such solution (λ, u) must coincide with (λ0 , u0 ). By Gidas, Ni and Nirenberg [GNN], both u and u0 are radially symmetric. It is readily checked that v(r) = u((λ0 /λ)1/2 r) satisfies

N −1 ′ ′ r v + λ0 r N −1 g(v) = 0,

v(0) = c, v ′ (0) = 0.

Since u0 satisfies the above equation with the same initial values, by uniqueness of solutions to the above initial value problem (see, for example, [NN], Proposition 2.35), we deduce v = u0 . In particular, v(r) > 0 for r ∈ [0, 1) and v(1) = 0. This implies that λ = λ0 and hence u = v = u0 . This finishes the proof.  We are now ready to prove the main result of this subsection. T HEOREM 4.5. Suppose Ω = B. Then there exists µ0 > 0 such that (4.4) has no positive solution for µ < µ0 , has exactly one positive solution for µ = µ0 and exactly two positive solutions for µ > µ0 . Moreover, the positive solution set {(µ, v)} of (4.4) forms a  Furthermore, if we denote the upper “⊂”-shaped smooth curve in the space R1 × C(B). and lower branches by   µ, v µ : µ0  µ < ∞

and



(µ, vµ ): µ0  µ < ∞ ,

respectively, then µ → v µ (x) is strictly increasing for any fixed |x| < 1, µ → vµ (0) is strictly decreasing, and lim v µ (x) = ∞ ∀|x| < 1,

µ→∞

lim vµ (0) = ξ,

µ→∞

ξ = 0 if N = 1, 2 and ξ > 0 if N > 2.

P ROOF. We first show that for µ large, (4.4) has at least one positive solution. Let φ ∈ C0∞ (B) satisfy φ
0 and maxB φ = 1. Let v be the unique solution of
v + φ = 0, v|∂B = 0; let v¯ be the unique solution of
v + µ = 0, v|∂B = 0. It is easy to check that for all large µ, v¯
v and they are upper and lower solutions to (4.4), respectively. Therefore there exists µ1 > 0 such that (4.4) has at least one positive solution provided that µ
µ1 . Now we can set  µ0 = inf µ > 0: (4.4) has a positive solution .

180

Y. Du

We claim that µ0 > 0. If not, there exists µi → 0 and vi such that
vi + µi e−1/vi = 0. Set v˜i = vi /vi ∞ . Then
v˜i + µi

e−1/vi v˜i = 0, vi

v˜i |∂B = 0.

As e−1/vi /vi is uniformly bounded, by standard elliptic regularity, v˜i W 2,p → 0. The Sobolev embedding theorem implies that v˜i → 0 uniformly. However, this is impossible as v˜i ∞ = 1. This contradiction implies that µ0 > 0. Again by standard elliptic regularity, we can further show that (4.4) with µ = µ0 has at least one positive solution, and we choose one of them and denote it as v0 . We claim that v0 must be a degenerate solution. If not, then by the implicit function theorem we can show that for µ less than but close to µ0 , (4.4) has at least one positive solution, and this contradicts the definition of µ0 . Since v0 is degenerate, our Lemma 4.3 implies that the solutions near (µ0 , v0 ) form a smooth curve which turns to the right in the (µ, v) space. We may denote the upper and lower branches by v µ and vµ respectively, where v µ (0) > vµ (0). As long as (µ, v µ ) and (µ, vµ ) are nondegenerate, the implicit function theorem ensures that we can continue to extend these two branches in the direction of increasing µ, and we still denote the extensions as v µ and vµ . This process of continuation towards larger values of µ for both branches may be stopped at some finite µ∗ by one of the following three possibilities: (i) v µn ∞ or vµn ∞ goes to infinity for some µn → µ∗ − 0; (ii) v µn ∞ or vµn ∞ goes to 0 for some µn → µ∗ − 0 (note that by the Harnack inequality, v µ and vµ can only lose positivity through vanishing on the entire domain); ∗ (iii) v µ or vµ∗ is a degenerate solution. However, (i) cannot occur since v µn and vµn are uniformly bounded by Lp estimates and Sobolev embedding theorem; (ii) cannot occur either as otherwise, denoting vn = v µn or vµn , 0 = λB 1



e−1/vn −µn vn



→ λB 1 > 0.

Finally, (iii) cannot occur. This is because, if, say, (µ, v µ ) becomes degenerate at µ = µ∗ , ∗ then Lemma 4.3 tells us that all the solutions near (µ∗ , v µ ) must lie to the right-hand side of it, which is a contradiction. Therefore we can always extend these two branches of solutions to µ = ∞. By Lemma 4.4, we see that µ → v µ (0) and µ → vµ (0) must be strictly monotone and µ v (0) > v0 (0) > vµ (0) for any µ ∈ (µ0 , ∞). Hence %  lim vµ (0) = ξ ∈ 0, v0 (0) ,

µ→∞

& lim v µ (0) = η ∈ v0 (0), ∞ .

µ→∞

Let us first show that η = ∞. In fact we show a little more than that. Let us denote wµ = Dµ v µ . By Lemma 4.3, we find that for µ > µ0 but close to µ0 , wµ > 0 in B.

Bifurcation and related topics

181

Define µ∗ := sup{µ > µ0 : wµ′ > 0 in B for all µ′ ∈ (µ0 , µ)}. We show that µ∗ = ∞. Differentiate the equation for v µ with respect to µ, we find that





 −
wµ = µf ′ v µ wµ + f v µ > µf ′ v µ wµ

∀x ∈ B,

wµ |∂B = 0.

Therefore wµ
, ≡ 0 for µ ∈ (µ0 , µ∗ ] and by applying the strong maximum principle to the above differential inequality with µ = µ∗ , we deduce wµ∗ > 0 in B, wµ′ ∗ (1) < 0.  norm, the above conclusion for wµ∗ implies Since µ → wµ is continuous in the C 2 (B) that wµ > 0 in B for all µ > µ∗ but close to µ∗ , a contradiction to the definition of µ∗ . Therefore wµ > 0 in B for all µ > µ0 . This implies that µ → v µ (r) is strictly increasing and v µ (r) > v0 (r). It follows that % % & µ µ µ& v µ (r) = (−
)−1 µe−1/v
(−
)−1 µ0 e−1/v0 = v0 (r) → ∞ µ0 µ0

as µ → ∞, for any r ∈ [0, 1). We next show that ξ > 0 when N > 2 and ξ = 0 when N = 1, 2. By Lemma 4.4, this would imply that all the positive solutions of (4.4) are contained in these two solution branches if we can show that there is no positive solution of (4.4) satisfying u(0)  ξ when ξ > 0. Let us note that when N > 2, ξ > 0 is a consequence of (4.8) in the proof of Lemma 4.2. When N = 2, we argue indirectly. Suppose that ξ > 0. Consider the initial value problem

′ ′ rz = −re−1/z ,

z(0) = ξ, z′ (0) = 0.

′ ′ rz = −re−1/z ,

z(0) = vµ∗ (0), z′ (0) = 0,

It is easily seen that z′ (r) < 0 for r ∈ (0, r0 ) as long as z is positive on (0, r0 ). If z remains positive on [0, ∞), then z(x) = z(|x|) = z(r) satisfies
z = −e−1/z < 0 on R2 and hence is a bounded subharmonic function on R2 . It is well known that in such a case, z ≡ const. Clearly this is impossible. Hence z has a first zero r0 > 0: z(r) > 0 in [0, r0 ) and z(r0 ) = 0. By continuous dependence of the solutions on the initial values, for µ∗ large, the unique solution z∗ of the initial value problem

has a first zero r ∗ close to r0 . But then v ∗ (r) = z∗ (r ∗ r) is a solution of (4.4) with v ∗ (0) = vµ∗ (0) but µ = (r ∗ )2 → r02 = µ∗ as µ∗ → ∞. This contradicts Lemma 4.4. Hence we must have ξ = 0. When N = 1, the proof is similar but simpler. The initial value problem now is changed to z′′ = −ze−1/z ,

z(0) = ξ, z′ (0) = 0,

and the existence of a first zero of z follows from z′′ < 0 on [0, ∞). Finally we show that if ξ > 0, then (4.4) has no positive solution with u(0)  ξ . In fact, if there is such a solution, then the argument we used above can be repeated to show

182

Y. Du

that there is a second smooth curve {(µ, u)} ˜ of positive solutions which is “⊂”-shaped and for (µ, v) ˜ on its upper branch, v(0) ˜ → ∞ as µ → ∞. But this implies that for any large number C > 0, there are at least two solutions v µ and v˜ with v µ (0) = v(0) ˜ = C, contradicting Lemma 4.4. The proof of Theorem 4.5 is complete. 

4.2. The perturbed Gelfand equation in dimensions 1 and 2 We will show that when Ω = B in dimensions 1 and 2, the positive solutions of the perturbed Gelfand equation are completely determined by (4.4). We use the equivalent form (4.3), namely −
u = µe−1/(ε+u)

in B,

u|∂B = 0,

(4.17)

where B = {x ∈ RN : |x| < 1}, N = 1, 2. Let us first observe the following simple relationship between (4.17) and (4.4). If (µ, v) is a positive solution of (4.4) with Ω = B, and v(0) > ε, then we can find a unique a ∈ (0, 1) such that v(a) = ε. Define u(x) = v(ax) − ε,

x ∈ B.

Clearly −
u = a 2 µf (u + ε),

u|∂B = 0.

That is, (a 2 µ, u) is a positive solution of (4.17). In dimensions 1 and 2, since ξ = 0 in Theorem 4.5, we find that for any ε > 0, we can obtain a positive solution of (4.17) from a positive solution of (4.4) through the above process. Moreover, by Lemma 4.4, we easily see that all the positive solutions of (4.17) can be obtained in this way. Therefore, the positive solutions of (4.17) are completely determined by (4.4). The following result, for N = 1, 2, is the counterpart of Lemma 4.2, but we will see in Remark 4.9 that this result is not true when 3  N  9. This indicates that, in this kind of conclusions, the nonlinearity and the space dimension play a very subtle role. L EMMA 4.6. If u is a degenerate positive solution of (4.17) and φ is a nontrivial solution to −
φ = µfε′ (u)φ,

φ|∂B = 0,

then φ does not change sign in B. P ROOF. Before starting the proof, let us remark that our proof only requires ε
0. Therefore, it is a simplification of the proof of Lemma 4.2 for the case N = 1, 2.

Bifurcation and related topics

183

By Gidas, Ni and Nirenberg [GNN], u is radially symmetric: u(x) = u(r), r = |x|; moreover, u′ (r) < 0 on (0, 1]. By Proposition 3.3 of [LN], φ is also radially symmetric, φ(x) = φ(r). Hence N −1 ′ φ + µfε′ (u)φ = 0 in [0, 1], r

φ ′′ +

φ ′ (0) = 0, φ(1) = 0.

We may assume φ(0) > 0. We make use of the test function v(r) = ru′ (r) + ξ instead of the usual v = ru′ + ξ u as used in the proof of Lemma 4.2, where ξ is a positive constant to be specified later. This choice of test function seems crucial. By a direct calculation, % & N −1 ′ v + µfε′ (u)v = µ ξfε′ (u) − 2fε (u) ≡ G(r), r &′ % N −1 ′ v φ − vφ ′ = G(r)r N −1 φ, r

v ′′ +

where

G(r) = µfε (u)g(r),

g(r) =

(4.18)

ξ − 2. (u + ε)2

Clearly, g(r) is strictly increasing in r. Now we suppose φ(r) changes sign in (0, 1), and want to deduce a contradiction from this. Let r0 ∈ (0, 1) be the first zero of φ(r): φ(r0 ) = 0 and φ(r) > 0 for r ∈ [0, r0 ). We choose ξ = −r0 u′ (r0 ) in v = ru′ + ξ . Since v ′ = −rµfε (u) + (2 − N )u′ < 0 ∀r ∈ (0, 1], we have v(r) > v(r0 ) = 0 on [0, r0 ) and v(r) < 0 on (r0 , 1]. We divide our considerations below into two cases: (i) g(r0 )  0 and (ii) g(r0 ) > 0. In case (i), using g(r) < g(r0 )  0 on [0, r0 ), we obtain the following contradiction by integrating (4.18) from 0 to r0 : 0>



0

r0

%

&r G(r)r N −1 φ dr = r N −1 v ′ φ − vφ ′ 00 = 0.

In case (ii), we consider the last zero of φ(r) before r = 1: r0  r 0 < 1, φ(r 0 ) = 0, φ(r) = 0 for r ∈ (r 0 , 1). We may assume that φ(r) > 0 on (r 0 , 1) (otherwise change the

184

Y. Du

sign of φ). Then φ ′ (r 0 ) > 0 > φ ′ (1). Now using g(r) > 0 and v(r)  0 on [r 0 , 1], we again deduce a contradiction:  1 %

&1 0< G(r)r N −1 φ(r) dr = r N −1 v ′ φ − vφ ′ r 0  0. r0



The proof is complete.

Using Lemma 4.6, we obtain a variant of Lemma 4.3, whose obvious proof we omit. L EMMA 4.7. Suppose that u0 is a degenerate positive solution of (4.17) with µ = µ0 .  lie on a Then all positive solutions (µ, u) of (4.17) that are near (µ0 , u0 ) in R × C(B) smooth curve represented by

 (µ, u) = µ0 + τ (s), u0 + sφ + z(s) with s small,

where z(0) = z′ (0) = 0, τ (0) = τ ′ (0) = 0 and φ is the positive eigenfunction given in Lemma 4.6. Moreover,  ′′ f (u0 )φ 3 dx ′′ τ (0) = −µ0 B ε . (4.19) B fε (u0 )φ dx We are now ready to prove the main result of this subsection.

T HEOREM 4.8. For all sufficiently small ε > 0, the positive solution curve {(µ, u)} of (4.17) is exactly S-shaped. There exist λ∗ε and Λ∗ε satisfying limε→0 λ∗ε = µ0 , limε→0 Λ∗ε = ∞, such that (4.17) has (i) a unique positive solution for µ ∈ (0, λ∗ε ) ∪ (Λ∗ε , ∞); (ii) exactly two positive solutions for µ = λ∗ε and µ = Λ∗ε ; (iii) exactly three positive solutions for µ ∈ (λ∗ε , Λ∗ε ). P ROOF. By Theorem 4.5, the positive solution curve of (4.4) with Ω = B is “⊂”-shaped with exactly one turning point at (µ0 , v0 ), where v0 = vµ0 = v µ0 . Denote ξ0 = v0 (0). Then for any ε ∈ (0, ξ0 ), we can find a unique µε ∈ (µ0 , ∞) such that vµε (0) = ε. By Theorem 4.5, we see that µε increases as ε decreases and µε → ∞ as ε → 0. For any µ ∈ [µ0 , µε ), we can find a unique aµ = aµ (ε) ∈ (0, 1) such that vµ (aµ ) = ε. Clearly, lim aµ (ε) = 1

ε→0

lim

µ→µε −0

for fixed µ
µ0 ,

aµ (ε) = 0 for fixed ε ∈ (0, ξ0 ).

(4.20)

Bifurcation and related topics

185

Now we define ηµ = (aµ )2 µ,

uµ (x) = vµ (aµ x) − ε,

x ∈ B,

and find that  Γε = (ηµ , uµ ): µ0  µ < µε

gives a piece of smooth solution curve to (4.17). Moreover, Γε connects (ηµ0 , uµ0 ) and (0, 0) (when µ → µε − 0). On the other hand, due to ε ∈ (0, ξ0 ), for any µ
µ0 , we can find a unique a µ = a µ (ε) ∈ (0, 1) satisfying v µ (a µ ) = ε, and define

2 ηµ = a µ µ,

 uµ (x) = v µ a µ x − ε,

x ∈ B,

we obtain another piece of smooth solution curve of (4.17) Γε =

 µ µ  η , u : µ0  µ < ∞ .

By Theorem 4.5, µ → a µ is strictly increasing and lim a µ = 1.

(4.21)

µ→∞

Therefore µ → ηµ is strictly increasing and {(µ, u(0)): (µ, u) ∈ Γ ε } is a monotone curve in R2 that connects (ηµ0 , uµ0 (0)) to (∞, ∞). Since

µ µ  η 0 , u 0 = (ηµ0 , uµ0 ),

we find that

Γ (ε) = Γ ε ∪ Γε gives a piecewise smooth (in fact smooth) curve for (4.17) connecting (0, 0) and (∞, ∞). By Lemma 4.4, we know it contains all the positive solutions of (4.17). We are going to find out the shape of this curve. Recall that f ′′ (u) > 0

for u ∈ (0, 1/2)

and f ′′ (u) < 0 for u ∈ (1/2, ∞).

We fix some ξ1 ∈ (0, 1/2) and suppose ε < ε1  1/2 − ξ1 . Then clearly f ′′ (u + ε) > 0 for u ∈ (0, ξ1 ).

186

Y. Du

Now we choose λξ1 > µ0 such that when µ
λξ1 .

vµ (0) < ξ1

By shrinking ε1 we may assume that λξ1 < µε for all ε ∈ (0, ε1 ). We can now divide Γε into two parts  Γε1 = (ηµ , vµ ): λξ1  µ < µε

We first analyze the shape of Γε1 . Define Λ∗ε =

sup

 and Γε2 = (ηµ , vµ ): µ0  µ  λξ1 .

ηµ .

µ∈[λξ1 ,µε )

By (4.20), one easily sees that there exists ε2 ∈ (0, ε1 ] such that when ε ∈ (0, ε2 ), Λ∗ε is achieved at some µ∗ ∈ (λξ1 , µε )

and

lim Λ∗ε = ∞.

ε→0

By the implicit function theorem, (ηµ∗ , uµ∗ ) must be a degenerate solution of (4.17). Then by Lemma 4.7, (4.19) and our choice of ξ1 , the solutions of (4.17) near (ηµ∗ , vµ∗ ) form a smooth curve that has a turn to the left. Therefore, we have an upper branch and a lower branch of positive solutions starting from this point, and both branches can be continued towards smaller values of µ. The lower branch can be continued to reach (0, 0), because (a) we cannot meet a degenerate solution in the way of continuation due to Lemma 4.7 and u(0) < ξ1 on Γε1 , and (b) the branch goes along Γε1 . For the same reason, the upper branch can be continued till it reaches (ηλξ1 , uλξ1 ). This implies that Γε1 is exactly “⊃”-shaped. Next we analyze the shape of Γε2 . It is more convenient for our discussion if we consider a bigger piece of solution curve Γε3 = Γε2 ∪

 µ µ  η , u : µ0  µ  λξ1 ,

which contains part of Γ ε . We observe that any (µ, u) ∈ Γε3 satisfies λ∗ε  µ  λξ1 ,

uλξ1 (0) − ε  u∞ = u(0)  uλξ1 (0) − ε,

(4.22)

where  λ∗ε = inf µ: (µ, u) ∈ Γε3 .

Since ηµ is increasing in µ, λ∗ε is achieved at some ηµ′ , µ′ ∈ [µ0 , λξ1 ). Therefore (λ∗ε , uµ′ ) must be a degenerate solution of (4.17). Clearly

2 λ∗ε  ηµ0 = aµ0 (ε) µ0 < µ0 .

Bifurcation and related topics

187

On the other hand, it is easy to see that aµ (ε) → 1 as ε → 0 uniformly for µ ∈ [µ0 , λξ1 ]. Hence  2 lim λ∗ε = lim min aµ (ε) µ: µ0  µ  λξ1 = µ0 .

ε→0

ε→0

We know from the above that Γε3 contains at least one degenerate solution (λ∗ε , uµ′ ). If we can show that there exists ε3 ∈ (0, ε2 ) such that whenever ε ∈ (0, ε3 ), any degenerate solution on Γε3 must make τ ′′ (0) > 0 in (4.19) of Lemma 4.7, then a continuation argument much as before shows Γε3 contains exactly one degenerate solution at µ = λ∗ε and the curve makes a turn to the right at this point. Hence Γε3 must be smooth and “⊂”-shaped. This tells us that the entire solution curve Γ (ε) is exactly S-shaped with two turning points at µ = λ∗ε and µ = Λ∗ε , respectively. Clearly, this would finish the proof of Theorem 4.8. It remains to show that there exists ε3 ∈ (0, ε2 ) such that any degenerate solution on Γε3 must make τ ′′ (0) > 0 in (4.19) of Lemma 4.7 as long as ε ∈ (0, ε3 ). We argue indirectly. Suppose for some εk → 0, we can find degenerate solutions (µk , uk ) ∈ Γε3k such that 

f ′′ (uk + εk )φk3 dx τk′′ (0) = −µk B k B f (u + εk )φk dx

 0,

where φk is the positive eigenfunction given in Lemma 4.7 when (µ, u) = (µk , uk ). We may assume that φk ∞ = 1. By (4.22), we may assume that µk → µ0 ∈ [µ0 , λξ1 ]. Equation (4.22) also implies that f (uk + εk )∞ is uniformly bounded. Therefore, by the equation for uk and a standard regularity and compactness argument, {uk } has a convergent subsequence in C 1 . We may assume uk → u0 in C 1 . Moreover, from

 −
φk = µk f ′ uk + εk φk ,

φk |∂B = 0,

we can use a similar regularity and compactness argument to obtain a C 1 convergent subsequence of φk . We may assume φk → φ 0 . Then we easily deduce

and

 −
u0 = µ0 f u0 ,

 u0 ∂B = 0, u0
0, u0 = 0,

 −
φ 0 = µ0 f ′ u0 φ 0 ,

  φ|∂B = 0, φ 0
0, φ 0 ∞ = 1.

This is to say (µ0 , u0 ) is a degenerate positive solution of (4.4) and φ 0 is the corresponding positive eigenfunction. By Theorem 4.5, (4.4) has a unique degenerate positive solution which is (µ0 , u0 ), and by Lemma 4.3 and (4.13),  ′′ f (u0 )φ 3 dx > 0. τ (0) = −µ0 B B f (u0 )φ dx ′′

188

Y. Du

Fig. 3. Bifurcation diagram for (4.17) with small ε > 0.

Therefore, we must have µk → µ0 , uk → u0 and φ 0 = φ (note that the positive eigenfunction is unique if it is normalized). Hence we have  ′′ k  ′′ 3 f (u0 )φ 3 dx ′′ k  B f (u + εk )φk dx B 0
τk (0) = −µ → −µ > 0. 0 k B f (u + ε)φk dx B f (u0 )φ dx 

This contradiction finishes our proof.

Theorem 4.8 is illustrated by the bifurcation diagram in Figure 3. 4.3. The perturbed Gelfand equation in higher dimensions In this subsection, we consider the perturbed Gelfand equation with space dimension N
3. For such N , by Theorem 4.5, any positive solution (µ, v) of (4.4) with Ω = B satisfies v(0) > ξ > 0,

where ξ = lim vµ (0). λ→∞

We can still use the transformation u(x) = v(ax) − ε, where v(a) = ε, to obtain a positive solution (a 2 µ, u) of (4.3) with Ω = B from a positive solution (µ, v) of (4.4). More precisely, we can define   Γ (ε) = Γ ε ∪ Γε , Γ ε = ηµ , uµ : µ0  µ < ∞ ,   Γε = ηµ , uµ : µ0  µ < ∞ , as in Section 4.2, except that now ηµ and uµ can be defined for all µ
µ0 , since ξ > 0 (we assume that ε < ξ ).

Bifurcation and related topics

189

However, unlike in the case N = 1 and 2, the solution curve Γ (ε) does not give all the positive solutions of (4.3) with Ω = B, since any (µ, u) ∈ Γ (ε) satisfies u(0) > ξ − ε, and it is well known (and can be easily shown) that for any fixed ε > 0, the entire positive solution branch {(µ, u)} of (4.3) with Ω = B has the point (0, 0) on its boundary: it approaches (0, 0) as µ → 0. Let us note that this implies that ηµ remains bounded for µ ∈ [µ0 , ∞). To better understand the entire positive solution branch for (4.3) with Ω = B, we need the help from the Gelfand equation (4.2) and some general bifurcation results; we will follow the approach in [Da1] and [Du2]. We first recall some abstract results. Suppose that (H1 ) X and Y are real Banach spaces and F : R2 × X → Y is a C k -map (k
2) which sends (ε, λ, x) ∈ R2 × X to F (ε, λ, x) ∈ Y ; (H2 ) Σ is a component of the set of solutions (λ, x) of F (0, λ, x) = 0 and Σ0 is a connected compact subset of Σ; (H3 ) For any (λ, x) ∈ Σ0 , Fx (0, λ, x) : X → Y is a Fredholm operator of index 0, and the mapping B(µ, h) := Fλ (0, λ, x) + Fx (0, λ, x) : R1 × X → Y is onto; in other words, for any (λ, x) ∈ Σ0 , either (i) Fx (0, λ, x) : X → Y has a continuous inverse, or (ii) dim N (Fx (0, λ, x)) = codim R(Fx (0, λ, x)) = 1 and Fλ (0, λ, x) ∈ / R(Fx (0, λ, x)). We want to know how Σ0 is perturbed to give a solution set Σε of F (ε, λ, x) = 0 when ε is small. We start with a local analysis. Suppose (λ0 , x0 ) ∈ Σ0 . Then either case (i) or case (ii) in (H3 ) happens. If case (i) happens, then it follows from the implicit function theorem that there exist a neighborhood V of x0 in X and a small number δ > 0 such that

 F −1 (0) ∩ (−δ, δ) × (λ0 − δ, λ0 + δ) × V   = ε, λ, x(λ, ε) : |ε|, |λ − λ0 | < δ ,

(4.23)

where (ε, λ) → x(ε, λ) is C k , and x(0, λ0 ) = x0 . If case (ii) happens, then by Theorem 1.3 and its proof (see [CR2]), there exist a small neighborhood W of (λ0 , x0 ) and a small δ > 0 such that

   F −1 (0) ∩ (−δ, δ) × W = ε, λ(s, ε), x(s, ε) : s, ε ∈ (−δ, δ) ,

(4.24)

where λ(s, ε) and x(s, ε) are C k with λ(0, 0) = λ0 , λs (0, 0) = 0, x(s, ε) = x0 + su0 + z(s, ε), τ (0, 0) = τs (0, 0) = 0. Here u0 spans N (Fx (0, λ0 , x0 )) and z(s, ε) belongs to a complement Z of span{u0 } in X. Since Σ0 is compact, by a finite covering argument, we find from the above discussion that Σ0 is a C k curve and for any small ε, the solutions of F (ε, λ, x) = 0 near Σ0 form a C k curve Σε ; see [Da1], Theorem 2, and [DLo1], Proposition A.2, for more details. If we change C k to “analytic” in (H1 ), then we can do the same in all the conclusions in the above discussions.

190

Y. Du

 Let us now see how the above discussions can be applied to (4.3). Set X = C02,α (B),  and F (ε, µ, u) =
u + µfε (u). Then from the proofs of Lemma 4.3 and Y = C α (B) Theorem 4.5 we know that the conditions (H1 )–(H3 ) are satisfied for positive λ and ε if we choose Σ0 to be any connected compact part of the global positive solution curve {(µ, u)} of (4.3) with Ω = B. Moreover, if we choose Σ0 to contain the unique turning point (µ0 , u0 ), then due to Lemma 4.3, (4.23) and (4.24), we find that for all small ε > 0, Σε contains exactly one turning point (µε , uε ). Moreover, it is easily seen that the upper branch in Σε contains part of Γ ε defined through the transformation at the beginning of this subsection. As in Section 4.3, we easily see from Theorem 4.5 that ηµ is a strictly increasing function. Thus this upper branch can be parameterized by µ and gives a unique positive solution for each fixed µ > µε . The lower branch can also be continued, and in fact, it can be continued till it reaches (0, 0); however, the shape of the lower branch is much more difficult to control. Let us note that if we have chosen Σ0 to be a rather large piece of the “⊂”-shaped positive solution curve of (4.4) (with Ω = B), then Σε with small ε > 0 is a large piece of “⊂”-shaped positive solution curve of (4.3) (with Ω = B); hence the shape of its lower branch is controlled till a very large µ. Let us denote the union of Σε and its well-controlled continued entire upper branch by Γε∗ . Then we know that Γε∗ is exactly “⊂”-shaped, its upper branch is unbounded and approaches (∞, ∞), and its lower branch contains points with large but finite µ. In order to better understand the rest of the global positive solution curve of (4.3), we now use the equivalent form (4.1), which can be viewed as a perturbation of the Gelfand equation (4.2). We should note that the curve Γε∗ for (4.3) is now transformed to a positive solution curve Γε∗ := {(λ, u): λ = µ/(ε 2 e1/ε ), u = v/ε 2 , (µ, v) ∈ Γε∗ } for (4.1) with Ω = B, its upper branch is still unbounded, but any (λ, u) belonging to its lower branch has λ small (since λ = µ/(ε 2 e1/ε ) and ε is small). Let us also note that any (λ, u) ∈ Γε∗ satisfies u(0) > (ξ − ε)/ε 2 → ∞ as ε → 0. We illustrate Γε∗ and Γε∗ in Figure 4. Let us consider the Gelfand equation (4.2) with Ω = B in more detail. When 3  N  9, we know from Proposition 4.1 and Figure 2 that its positive solution curve has infinitely

Fig. 4. Bifurcation diagrams for (4.1) and (4.3) with Ω = B and small ε > 0.

Bifurcation and related topics

191

many turning points. Let us denote them by Tk , k = 1, 2, . . . , where the Tk ’s are ordered so that when we go from T0 := (0, 0) along the solution curve we meet T1 first, then T2 , T3 , etc. It is known (see [Da1] and [NS], Section 2) that if (λ, u) lies on the open arc from Tk to Tk+1 , then u is nondegenerate; and if (λ, u) = Tk with k
1, then u is degenerate and the nontrivial solution φ to −
φ = λeu φ,

φ|∂B = 0,

is radially symmetric, φ(x) = φ(r); moreover, φ(r) changes sign over [0, 1) exactly k − 1 times, and is unique up to a scalar multiple, and 

B

eu φ 3 dx = 0.

(4.25)

 Y = C α (B)  and F (ε, λ, u) =
u + These facts imply that if we take X = C02,α (B), u/(1+εu) , then the conditions (H1 )–(H3 ) are satisfied with λ, ε > 0. Moreover, case (ii) λe in (H3 ) occurs exactly when (λ, u) ∈ {Tk : k
1}. By (4.24) and (4.25), we deduce that τss (0, 0) = 0 when (λ, u) = Tk , k
1. Hence we can apply the above abstract results to conclude the following: Let Σ0 be a connected compact part of the positive solution curve {(λ, u)} of (4.2) with Ω = B, and suppose that Σ0 contains the turning points T1 , . . . , Tk . Then for all small ε > 0, the positive solution curve of (4.1) with Ω = B has a part Σε , which is close to Σ0 and contains exactly k turning points T1ε , . . . , Tkε , where Tiε → Ti as ε → 0, for each i = 1, . . . , k. Close to T0 = (0, 0), we can apply the implicity function theorem to (4.1) to conclude that there is a unique positive solution for each λ > 0 small. Therefore we can extend Σε towards (0, 0) so that T0 is an end point of Σε , provided that Σ0 has been chosen properly. R EMARK 4.9. It is easy to show that, at Ti , 2  i  k, the linearized problem of (4.1) (with Ω = B and small ε > 0) has a radially symmetric nontrivial solution φ which changes sign exactly i − 1 times over [0, 1). Hence Lemma 4.6 does not hold when 3  N  9. If N
10, then it is known (see [Da1]) that the positive solution curve of (4.2) with Ω = B contains no turning point and all the positive solutions are nondegenerate. Therefore a similar consideration to the above yields the following conclusion: Given any connected compact part Σ0 of the positive solution curve {(λ, u)} of (4.2) with Ω = B, for all small ε > 0, the positive solution curve of (4.1) with Ω = B has a part Σε , which is close to Σ0 and contains no turning points. Moreover, Σε can be extended to reach (0, 0) and the extended Σε still does not contain any turning point. Combined with Γε , we can now illustrate the well-understood parts of the positive solution curve of (4.1) in Figure 5. The missing part of the positive solution curve for (4.1) can be further described by using the bifurcation curve S0 := {(λ, u′ (1))} of (4.2) with Ω = B; this curve is completely understood in [JL], and Dancer [Da1] used S0 to show that the missing part is roughly the

192

Y. Du

Fig. 5. Bifurcation diagrams for (4.1) with Ω = B and small ε > 0.

bottom part in Figure 5 reflected about the horizontal line passing through the top point of the bottom part. However, on this reflected part, the exact number of turning points is not known. We refer to [Da1], Figure 1, and [Du2], Figure 4, for bifurcation diagrams illustrating the entire positive solution curve.

4.4. Further remarks and related results The above results for the perturbed Gelfand equation (4.1) with Ω = B can be extended to the following more general problem −
u = λ(1 + εu)m eu/(1+εu) ,

u|∂B = 0,

(4.26)

where 0  m < 1; see [Du2] for details. The case m
1 was also discussed in [Du2], and the shape of the global positive solution curve for this case is very different. Equation (4.26) arises from combustion theory. In catalysis theory, there is an equation closely related to (4.1); under some simple changes of variables, it reduces essentially to −
u = λ(1 − εu)p eu/(1+εu) ,

u|∂B = 0,

(4.27)

where p is a nonnegative integer (see, e.g., Aris [Ar], Vol. 1, Chapter 4). Let us note that when p = 0, (4.27) reduces to the perturbed Gelfand equation (4.1), and when ε = 0, (4.27) reduces to the Gelfand equation (4.2). For (4.27), it has been conjectured that for any nonnegative integer p, the positive solution set {(λ, u)} is S-shaped provided ε > 0 is small and the dimension N = 1 or 2. The conjecture was proved to be true for N = 1 by Hastings and McLeod [HM]. For N = 2 and p > 0, the conjecture is proved in [Da1] except for a small λ-range (with values close to 0); the higher-dimension case was also considered in [Da1].

Bifurcation and related topics

193

In a series of papers, Korman, Li, Ouyang and Shi developed various useful techniques for proving exact multiplicity results for positive solutions to equations of the form −
u = λf (u),

u|∂B = 0,

with several classes of nonlinearities; see [OS1,OS2] and the references therein. Our arguments in Section 4.1 basically follow their strategy. See also [Tang] for a more recent result along this line. The abstract results discussed and used in Section 4.3 was further developed by Shi [S].

5. Nodal properties and global bifurcation To a large extent, the approach in the previous section relies on the analysis of nodal properties of the solutions. In this section, we discuss some other aspects of the nodal properties and their applications in bifurcation theory. Due to the uniqueness of initial value problems in ODEs, it is well known that the nodal properties of the solutions to a Sturm–Liouville problem for a second-order ordinary differential equation do not change when one moves along a global bifurcation branch of the solutions. This fact can be used to show that global bifurcation branches bifurcating from different eigenvalues do not meet each other and hence yielding multiple existence results; see [Ra], Theorem 2.3. This idea has been extended to elliptic partial differential equations over various two-dimensional domains with certain symmetries, in particular rectangles, by Kielhofer and his collaborators; see [K], Section III.6. In this section we look at two examples where nodal properties of solutions are used for systems of elliptic equations; here we consider radial solutions over circular domains. Let us note that, even for onedimensional domains (i.e., intervals), the analysis of the nodal properties for systems is not trivial. It is an interesting problem to see how the techniques in [K] for equations over twodimensional domains such as rectangles can be extended to systems (see [HK] for some efforts in this direction). Likewise, the techniques of our Section 4 may have applications to radial solutions for elliptic systems. Our first example examines a competition system and is taken from Nakashima [N]; our second example studies a predator–prey system and follows Dancer, López-Gómez and Ortega [DLO].

5.1. Global bifurcation of the competition system We consider the competition system with constant coefficients and Neumann boundary conditions   −
u = λu(α − u − βv), −
v = λv(γ − v − δu),  ∂ν u|∂Ω = ∂ν v|∂Ω = 0,

(5.1)

194

Y. Du

 = {x ∈ RN : R0  |x|  R1 }, 0  R0 < R1 , where Ω is either a ball or an annulus: Ω N
1. We assume throughout this subsection that the positive coefficients satisfy δ −1
0, the two semitrivial solutions (α, 0) and (0, γ ) are linearly stable, and (5.1) has a unique constant positive solution

∗ ∗ u ,v =



βγ − α αδ − γ , , βδ − 1 βδ − 1

which is linearly unstable. We are interested in the existence of other positive solutions of (5.1). We will make use of the nodal properties of radially symmetric positive solutions and global bifurcation theory to address this problem. L EMMA 5.1. Suppose that (u(x), v(x)) = (u(r), v(r)) is a radial, nonconstant positive solution of (5.1), and v ′ (r)
0 for r ∈ (r∗ , r ∗ ) ⊂ (R0 , R1 ), and u′ (r∗ ), u′ (r ∗ )  0. Then u′ (r) < 0 in (r∗ , r ∗ ). Similarly, if v ′ (r)  0 for r ∈ (r∗ , r ∗ ) ⊂ (R0 , R1 ), and u′ (r∗ ), u′ (r ∗ )
0. Then u′ (r) > 0 in (r∗ , r ∗ ). Note that the roles of u and v in this lemma can be interchanged. P ROOF OF L EMMA 5.1. We only consider the first case; the proof for the second is parallel. Denote f (u, v) = λ(α − u − βv) and g(u, v) = λ(γ − v − δu). We have  N −1 ′ ′′   u + r u + uf (u, v) = 0, v ′′ + N r−1 v ′ + vg(u, v) = 0,   ′ u (R0 ) = u′ (R1 ) = v ′ (R0 ) = v ′ (R1 ) = 0.

Let φ(r) = u′ (r)/u(r). We easily find φ′ + φ2 +

N −1 φ + f (u, v) = 0. r

Differentiating this identity we obtain   N −1 ′ N −1 φ + ufu (u, v) − φ + fv (u, v)v ′ = 0. φ ′′ + 2φ + r r2 Since fu (u, v) < 0, fv (u, v) < 0 and v ′
0 in (r∗ , r ∗ ), we find that, for r ∈ (r∗ , r ∗ ), c(r) := ufu (u, v) −

N −1 1), we deduce that (u, v) ≡ (ξ, η) in (R0 , R1 ), a contradiction to our assumption. This finishes our proof of Step 1. S TEP 2. If r1 < r2 < r3 are three consecutive zeros of u′ , then u′ (r) changes sign as r crosses r2 ; similarly, if r1 < r2 < r3 are three consecutive zeros of v ′ , then v ′ (r) changes sign as r crosses r2 . Again we only prove the first conclusion; the proof for the second is parallel. Suppose that u′ (r) does not change sign at r2 . We assume for definiteness that u′ (r) > 0 on both (r1 , r2 ) and (r2 , r3 ). We claim that there exists s1 ∈ (r1 , r2 ) such that v ′ (s1 ) < 0. Otherwise, we can apply Lemma 5.1 over (r1 , r2 ) to deduce u′ < 0 in (r1 , r2 ). Similarly there exists s3 ∈ (r2 , r3 ) such that v ′ (s3 ) < 0. We now claim that there exists s2 ∈ (s1 , s3 ) such that v ′ (s2 ) > 0; otherwise, we can apply Lemma 5.1 over (s1 , s3 ) to deduce u′ > 0 over (s1 , s3 ), contradicting the assumption that u′ (r2 ) = 0. Let (s∗ , s ∗ ) be the largest interval containing s2 in (s1 , s3 ) such that v ′ > 0 in (s∗ , s ∗ ) and v ′ (s∗ ) = v ′ (s ∗ ) = 0. We now apply Lemma 5.1 over (s∗ , s ∗ ), where u′
0, and deduce that v ′ < 0 in (s∗ , s ∗ ). This contradiction shows that u′ (r) changes sign at r2 . This proves Step 2. S TEP 3. If R0 = s0 < s1 < · · · < sm = R1 and R0 = t0 < t1 < · · · < tn = R1 are the zeros of u′ and v ′ , respectively, then m = n and for all possible k, sk−1 < tk < sk+1 ,

tk−1 < sk < tk+1 ,

and the sign of u′ in (sk , sk+1 ) is opposite to that of v ′ in (tk , tk+1 ). For definiteness, we assume that u′ (r) > 0 over (s0 , s1 ). We show that v ′ (r) < 0 over (t0 , t1 ). Suppose for contradiction that v ′ (r) > 0 in (t0 , t1 ). If s1  t1 , then we apply Lemma 5.1 over (s0 , s1 ) and deduce the contradiction that u′ (r) < 0 in (s0 , s1 ). If s1 > t1 , then applying Lemma 5.1 over (t0 , t1 ) we obtain v ′ (r) < 0 over (t0 , t1 ), again a contradiction. Therefore we must have v ′ < 0 over (t0 , t1 ). By Step 2 we now have (−1)k u′ (r) > 0 ∀r ∈ (sk , sk+1 ),

(−1)l v ′ (r) < 0

∀r ∈ (tl , tl+1 ).

(5.4)

We must have s1 < t2 for if s1
t2 then we can apply Lemma 5.1 over (t1 , t2 ) to deduce v ′ < 0 on (t1 , t2 ). We can similarly prove that s2 < t3 ; for otherwise we can apply Lemma 5.1 over (t2 , t3 ) ⊂ (s1 , s2 ) to deduce v ′ > 0 in (t2 , t3 ). Continuing in this fashion, we can show that sk < tk+1 for all possible k, namely k = 1, 2, . . . , min{m, n − 1}. If n > m, we already have a contradiction when we take k = m: R1 = sm < tm+1 . Hence we must have n  m. In a similar fashion, we can show that tk < sk+1 for k = 1, 2, . . . , min{n, m − 1} and m  n. Therefore we must have m = n. This finishes the proof of Step 3.

Bifurcation and related topics

197

S TEP 4. u′′ (sk ) = 0, v ′′ (tk ) = 0 for k = 0, 1, . . . , m. We will show that u′′ (sk ) = 0; the proof for v ′′ (tk ) = 0 is similar. For definiteness, we assume that (5.4) holds. As in the proof of Lemma 5.1, we have, for φ = u′ /u,  N −1 ′ N −1 φ = −fv (u, v)v ′ . φ + ufu (u, v) − φ + 2φ + r r2 ′′



(5.5)

If sk = s0 = R0 , we find from (5.5) and (5.4) that

  N −1 ′ N −1 φ ′′ + 2φ + φ + ufu (u, v) − φ < 0 ∀r ∈ (R0 , t1 ). r r2

Since φ(r) > 0 in (R0 , s1 ) and (2φ + (N − 1)/r) + (r − R0 )(ufu (u, v) − (N − 1)/r 2 ) is always bounded from below (even if R0 = 0), we can apply Theorem 4 on page 7 of [PW] to conclude that φ ′ (R0 ) > 0, which implies u′′ (R0 ) > 0. If sk = sm = R1 , the proof is similar. Next we consider the case that sk ∈ (R0 , R1 ). For definiteness we assume that u′ > 0 in (sk , sk+1 ). If tk = sk , then the proof is similar to the case that sk = R0 . Suppose now tk = sk . If tk > sk , then by what is proved in Steps 2 and 3 we must have v ′ > 0 in (tk−1 , sk ], and we obtain from (5.5) that   N −1 ′ N −1 φ > 0 ∀r ∈ (tk−1 , sk ). φ + ufu (u, v) − φ ′′ + 2φ + r r2 Since φ < 0 in (sk−1 , sk ), we can apply [PW], page 7, Theorem 4, on some (sk − ε, sk ) to obtain φ ′ (sk ) > 0 and hence u′′ (sk ) > 0. If tk < sk , then v ′ < 0 in [sk , tk+1 ) and   N −1 N −1 ′ φ + ufu (u, v) − φ < 0 ∀r ∈ [sk , tk+1 ). φ ′′ + 2φ + r r2

Since φ > 0 in (sk , sk+1 ), we obtain from [PW] that φ ′ (sk ) > 0 and hence u′′ (sk ) > 0. This finishes the proof for Step 4 and hence completes the proof of Theorem 5.2.  Following [N] we say a function pair (u, v) ∈ C 2 ([R0 , R1 ])2 has mode m if u′ (R0 ) = ′ ′ ′ ′ 1 ) = 0, v (R0 ) = v (R1 ) = 0 and (u , v ) has the properties (i)–(iii) in Theorem 5.2; we denote

u′ (R



2 Sm := (u, v) ∈ C 2 [R0 , R1 ] : (u, v) has mode m .

Clearly each Sm , m = 1, 2, . . . , is an open set in C 2 ([R0 , R1 ]). We will use a global bifurcation argument to show that for any m
1, (5.1) has a radial positive solution (u, v) ∈ Sm , provided that λ is large enough. Let us now regard α, β, γ , δ as fixed positive parameters satisfying (5.2) and regard λ as a bifurcation parameter. Clearly (λ, u∗ , v ∗ ) solves (5.1) for any λ. We will call this solution

198

Y. Du

the trivial solution, and look for nontrivial positive solutions. We will restrict to the class of radial solutions and use a bifurcation approach. The linearized eigenvalue problem of (5.1) at (u∗ , v ∗ ) in the class of radial functions is given by 

∗  ′′ −1 ′ ∗   φ + (N − 1)r φ − λ u φ + βu ψ = −µφ,

 (5.6) ψ ′′ + (N − 1)r −1 ψ ′ − λ δv ∗ φ + v ∗ ψ = −µψ,   ′ ′ φ (Ri ) = ψ (Ri ) = 0, i = 0, 1. Denote

A=



−u∗ −δv ∗

−βu∗ −v ∗



.

Due to (5.2), it is easily seen that A has a positive eigenvalue κ + and a negative eigenvalue κ − . Therefore there exists a matrix P such that  −  − p1 p1+ 0 κ , where P = . P −1 AP = 0 κ+ p2− p2+ Define 

Φ Ψ



= P −1



φ ψ



;

then (5.6) becomes  ′′ −1 ′ −   Φ + (N − 1)r Φ + λκ Φ = −µΦ, ′′ −1 ′ Ψ + (N − 1)r Ψ + λκ + Ψ = −µΨ,   ′ Φ (Ri ) = Ψ ′ (Ri ) = 0, i = 0, 1.

(5.7)

+ The eigenvalues of (5.7) are easily found to be µ− n , µn , n = 0, 1, . . . , where 0 ± µ± n = µn − λκ ,

and 0 = µ00 < µ01 < µ02 < · · · are the eigenvalues of the problem −φ ′′ − (N − 1)r −1 φ ′ = µφ,

φ ′ (R0 ) = φ ′ (R1 ) = 0.

(5.8)

Hence, zero is an eigenvalue of (5.7) exactly when λ = λ± n :=

µ0n , κ±

n = 0, 1, 2, . . . ,

+ with the corresponding eigenfunction (Φn , 0) for λ− n , and (0, Φn ) for λn , where Φn is the ′ nth eigenfunction of (5.8). It is well known that Φn has exactly n − 1 zeros in (R0 , R1 ).

Bifurcation and related topics

199

Returning to (5.6), we find that zero is an eigenvalue of (5.6) exactly when λ = λ± n ,n = 0, 1, 2, . . . , and the corresponding eigenfunctions are given by

± ± ±  φn , ψn = p1 Φn , p2± Φn ,

where p1± , p2± come from P and by a simple analysis one finds that p1+ p2+ < 0, p1− p2− < 0. Therefore (φn± , ψn± ) ∈ Sn+1 . The above discussions enable us to apply both Theorems 1.1 and 1.2 to a suitable abstract version of (5.1) in the space of radial functions to conclude the following theorem. T HEOREM 5.3. Let S denote the closure of the radial nonconstant positive solutions (λ, u, v) of (5.1) in the space R1 × C 2 ([R0 , R1 ])2 . Then for each n
1, S contains a maximal subcontinuum Sn such that ∗ ∗ (i) Sn meets (λ+ n , u , v ); (ii) Sn either is unbounded or meets (λˆ , u∗ , v ∗ ) for some λˆ ∈ {λ± m : m = 0, 1, 2, . . . } \ + {λn }; ∗ ∗ 1 2 2 (iii) in a small neighborhood On± of (λ± n , u , v ) in R × C ([R0 , R1 ]) , all the radial solutions of (5.1) lie on a smooth curve (λ(s), u(s), v(s)), |s| < s0 , where 



 u(s), v(s) = u∗ , v ∗ + s φn± , ψn± = o(s). λ(s) = λ± n + o(s),

Let us analyze Sn in more detail. By a well-known result of Conway, Hoff and Smoller [CHS], there exists λ0 > 0 small enough such that for λ ∈ (0, λ0 ], (5.1) has only constant nonnegative solutions. We may assume that λ0 ∈ (0, λ+ 1 ). ∗ , v ∗ )} satisfies λ > λ , (u, v) ∈ S We claim that any (λ, u, v) ∈ Sn \ {(λ+ , u 0 n+1 and n u(r) > 0, v(r) > 0 in [R0 , R1 ]. In other words,

   ∗ ∗ Sn \ λ+ n ,u ,v  ⊂ ∆n := (λ, u, v): λ > λ0 , (u, v) ∈ Sn+1 , u > 0, v > 0 . (5.9)

Otherwise, by the connectedness of Sn and conclusions (i) and (iii) in Theorem 5.3, there ∗ ∗ exists some (λ∗ , u∗ , v∗ ) ∈ (Sn \ {(λ+ n , u , v )}) ∩ ∂∆n . Hence (u∗ , v∗ ) is a nonnegative solution of (5.1) with λ = λ∗ . If it is not a positive solution, then by the strong maximum principle, (u, v) has at least one component identically zero. This implies that (u, v) ∈ {(0, 0), (α, 0), (0, γ )}. As bifurcation to positive solutions from these solutions is possible only at λ = 0 (due to (5.2)), we find this is a contradiction. Therefore (u∗ , v∗ ) has to be a positive solution. If (u∗ , v∗ ) = (u∗ , v ∗ ), then we can find (λk , uk , vk ) ∈ ∆n that converges to (λ∗ , u∗ , v ∗ ) in R1 × C 2 ([R0 , R1 ])2 . By conclusion (iii) in Theorem 5.3, this is possible only if − − λ∗ = λ± n . Since λ∗
λ0 > 0 and λn < 0, λ∗ = λn is impossible. Since (λ∗ , u∗ , v∗ ) ∈ + ∗ ∗ + (Sn \ {(λn , u , v )}), λ∗ = λn is also impossible. Therefore (u∗ , v∗ ) must be a nonconstant positive solution of (5.1), and by the definition of λ0 , we know that λ∗ > λ0 . Moreover, by Theorem 5.2, (u∗ , v∗ ) ∈ Sm+1 for some m
0. Since (u∗ , v∗ ) can be approached in the C 2 norm by functions in Sn+1 , we necessarily have m = n. But then (λ∗ , u∗ , v∗ ) belongs to the interior of ∆n . This contradiction proves that (5.9) is true.

200

Y. Du

By (5.9) and conclusion (iii) in Theorem 5.3, we know that the second alternative in (ii) of Theorem 5.3 never occurs. Hence Sn is unbounded. A simple comparison argument shows that every nonnegative solution of (5.1) satisfies u  α and v  γ . Therefore by standard elliptic estimate we know that u and v have bounded C 2 norms as long as λ > 0 stays bounded. This implies that Sn can become unbounded only through λ → ∞. To summarize, we have the following theorem. ∗ ∗ T HEOREM 5.4. For each n
1, (λ, u, v) ∈ Sn \ {(λ+ n , u , v )} implies that (u, v) is a nonconstant positive radial solution of (5.1); moreover, (u, v) belongs to Sn+1 , and the projection of Sn on R1 contains (λ+ n , ∞), and is contained in (λ0 , ∞) with some λ0 > 0. Hence for each λ > λ+ , (5.1) has at least m nonconstant positive radial solutions. m

R EMARK 5.5. 1. The conclusion of Theorem 5.4 can be strengthened. In fact, each Sn can be decomposed into two unbounded subbranches Sn+ and Sn− satisfying Sn+ ∩ Sn− = ∗ ∗ + {(λ+ n , u , v )}. Hence (5.1) has at least 2m nonconstant positive solutions for λ > λm .  one easily 2. By considering the linearization of (5.1) at (u∗ , v ∗ ) in the space C 2 (Ω), sees that the linearized eigenvalue problem has eigenvalues corresponding to nonradially symmetric eigenfunctions. This fact can be used to construct examples where (5.1) has many nonradial positive solutions bifurcating from (u∗ , v ∗ ). 3. For large β and δ, one can also use the method in [DD1] to construct nonradially symmetric positive solutions  for (5.1). 4. It is unclear whether ∞ n=1 Sn contains “all” the radial nonconstant positive solutions of (5.1) (with λ > 0). 5. In Kan-on [Ka], it is shown that if α = γ and β = δ in (5.1), and if the space dimension is one, then each Sn is exactly “⊂”-shaped, and ∞ n=1 Sn contains “all” the nonconstant positive solutions of (5.1) (with λ > 0). 6. If Ω is a ball (in general, if it is a convex set), then any nonconstant positive solution of (5.1) is unstable; see [KW].

5.2. Uniqueness results for the predator–prey system We consider the Lotka–Volterra predator–prey system with constant coefficients and homogeneous Dirichlet boundary conditions  2   −
u = λu − u − cuv, −
v = µv − v 2 + duv,   u|∂Ω = v|∂Ω = 0,

(5.10)

where Ω is a bounded smooth domain in RN . It has long been conjectured that (5.10) has at most one positive solution, but the conjecture remains open except when the space dimension is one, namely when Ω is an interval. In contrast, the existence problem for (5.10) is completely understood. In the following discussion, we will use some notations of Section 3. If λ > λΩ 1 , then (θλ , 0) solves (5.10).

Bifurcation and related topics

201

Similarly, (0, θµ ) solves (5.10) when µ > λΩ 1 . Suppose that (u0 , v0 ) is a positive solution to (5.10). Then from the equation for u0 we obtain Ω λ = λΩ 1 (u0 + cv0 ) > λ1 .

Similarly, from the equation for v0 it follows µ = λΩ 1 (v0 − du0 ). A simple comparison argument shows that u0  θλ and v0
θµ ; here we use the convention that θµ = 0 if µ  λΩ 1 . Therefore Ω µ
λΩ 1 (v0 − dθλ ) > λ1 (−dθλ ) := µ0 .

Similarly, Ω λ
λΩ 1 (u0 + cθµ ) > λ1 (cθµ ). Ω 0 Suppose λ > λΩ 1 and let µ > λ1 be uniquely determined by

λ = λΩ 1 (cθµ0 ). We find from the above discussions that a necessary condition for (5.10) to possess a positive solution is that λ > λΩ 1 ,

µ0 < µ < µ0 .

(5.11)

We show next that (5.11) is also sufficient for (5.10) to have a positive solution. To this end, we fix the positive coefficients λ, c, d and regard µ as a bifurcation parameter. Under (5.11), we can apply the local and global bifurcation analysis much as for the competition model in Section 3.1, to conclude that there is a global bifurcation branch Σ = {(µ, u, v)} of positive solutions to (5.10), which bifurcates from the semitrivial branch {(µ, θλ , 0)} at (µ0 , θλ , 0), and joins the other semitrivial branch {(µ, 0, θµ )} at (µ0 , 0, θµ0 ). Therefore, for any µ ∈ (µ0 , µ0 ), (5.10) has at least one positive solution (u, v) such that (µ, u, v) ∈ Σ. In the following, we will make use of the nodal properties of certain differential systems to show that if Ω is an annulus,  Ω = x ∈ RN : R1 < |x| < R2 ,

0 < R1 < R2 ,

(5.12)

then the global bifurcation branch Σ consists of radially symmetric positive solutions only, and it is a smooth curve, and for each µ ∈ (µ0 , µ0 ), there is exactly one positive solution (u, v) of (5.10) such that (µ, u, v) ∈ Σ . Suppose that (5.12) holds. Let us first observe that, by uniqueness, θλ and θµ are radially symmetric, and if we apply the local and global bifurcation analysis in the space

202

Y. Du

of radially symmetric functions, we obtain a global branch of radially symmetric positive solutions of (5.10), which we denote by Σr , and Σr connects the semitrivial solution branches at (µ0 , θλ , 0) and (µ0 , 0, θµ0 ), respectively. We will show that Σ = Σr . We do this by considering the linearization of (5.10) at an arbitrary radially symmetric positive solution (u0 (x), v0 (x)) = (u0 (r), v0 (r))   −
φ = (λ − 2u0 − cv0 )φ − cu0 ψ, −
ψ = (µ − 2v0 + du0 )ψ + dv0 φ,  φ|∂Ω = ψ|∂Ω = 0.

(5.13)

 2 solves (5.13). Then T HEOREM 5.6. Suppose that (5.12) holds and (φ, ψ) ∈ C 2 (Ω) (φ, ψ) = (0, 0). P ROOF. We assume that N
2; the case N = 1 can be proved by similar arguments and is simpler. Arguing indirectly, we assume that (5.13) has a solution (φ, ψ) = (0, 0). We are looking for a contradiction. S TEP 1. Reduction to ODE systems by harmonic polynomials. We first use harmonic polynomials to reduce (5.13) to a sequence of ODE systems. (This is not needed if N = 1.) Let Hn denote the space of homogeneous and harmonic polynomials of degree n in N variables. It is well known that the restriction of these polynomials to the unit sphere S N −1 are the eigenfunctions of the Laplace–Beltrami operator
S N−1 corresponding to the eigenvalue λn = −n(n + N − 2), n = 0, 1, 2, . . . . Moreover, the following orthogonal decomposition holds

 - n L2 S N −1 = H .

(5.14)

n
0

Given Pn ∈ Hn define f (r) =



S N−1

φ(rξ )Pn (ξ ) dσ (ξ ),

g(r) =



S N−1

ψ(rξ )Pn (ξ ) dσ (ξ ).

Then (f, g) ∈ C 2 ([R1 , R2 ])2 , and f (R1 ) = f (R2 ) = 0, g(R1 ) = g(R2 ) = 0. Moreover, since the Laplace operator in RN can be expressed as
=
r + r −2
S N−1 , where r = |x|,


r = r 1−N ∂r r N −1 ∂r ,

Bifurcation and related topics

203

if we multiply the differential equations in (5.13) by Pn (ξ ) and integrate on S N −1 , it results  −
r f + n(n + N − 2)r −2 f        = (λ − 2u0 − cv0 )f − cu0 g, r ∈ (R1 , R2 ), −
r g + n(n + N − 2)r −2 g     = (µ − 2v0 + du0 )g + dv0 f, r ∈ (R1 , R2 ),    f (Ri ) = g(Ri ) = 0, i = 1, 2.

(5.15)

S TEP 2. There exists some Pn ∈ Hn , n
0, such that problem (5.15) has a solution (f, g) = (0, 0). Otherwise, for all Pn ∈ Hn , n
0, (5.15) has only the solution (0, 0). Hence for each r ∈ (R1 , R2 ), φ(rξ ) is orthogonal in L2 (S N −1 ) to every Hn , n
0. It follows from (5.14) that φ(rξ ) = 0. Hence φ ≡ 0 in Ω. Similarly ψ ≡ 0. But this contradicts our assumption. S TEP 3. The zeros of f and g in [R1 , R2 ] are isolated. We only prove the conclusion for f ; that for g is similar. Suppose for contradiction that r0 ∈ [R1 , R2 ] is an accumulating point of zeros of f in [R1 , R2 ]. Then we easily see that f (r0 ) = f ′ (r0 ) = 0. If g(r0 ) = g ′ (r0 ) = 0, then we deduce (f, g) ≡ (0, 0) by uniqueness of the initial value problem for (5.15). Hence either g(r0 ) = 0 or g ′ (r0 ) = 0. Therefore we can find an open interval I ⊂ [R1 , R2 ] with r0 as an end point and a sequence of zeros {rk } of f such that {rk } ⊂ I, rk → r0 and g(r) = 0 for all r ∈ I . From the equation for u0 we find λΩ 1 (−λ + u0 + cv0 ) = 0. It follows that

 −2 λΩ > 0. 1 −λ + 2u0 + cv0 + n(n + N − 2)r

If we denote Ω1 = {x ∈ Ω: |x| ∈ I1 }, where I1 denotes the interval with end points r0 and r1 , we obtain 

−2 1 λΩ 1 −λ + 2u0 + cv0 + n(n + N − 2)r

 −2 > 0. > λΩ 1 −λ + 2u0 + cv0 + n(n + N − 2)r

This implies that the strong maximum principle is satisfied by the operator

 L1 := −
− λ + 2u0 + cv0 + n(n + N − 2)r −2

over Ω1 : L1 u
0 in Ω1 and u
0 on ∂Ω1 imply u > 0 or u ≡ 0 in Ω1 . If we identify f (r) with f (|x|), then from (5.15) we obtain L1 f = −cu0 g and f |∂Ω1 = 0. Therefore by the strong maximum principle we have either f ≡ 0 in Ω1 or f does not vanish and has the opposite sign to g in Ω1 . If f is identically 0 then by (5.15) we must have g identically zero over Ω1 and hence by uniqueness of the initial value problem for (5.15) we deduce (f, g) = (0, 0) over Ω; a contradiction. If f does not vanish

204

Y. Du

in Ω1 , then we arrive a contradiction with the fact that I1 contains infinitely many zeros of f (r). Therefore the zeros of f in [R1 , R2 ] must be isolated. S TEP 4. Let R1 < ξ1 < · · · < ξp = R2 be the finite sequence of zeros of f where f changes sign, and assume that f
0 in (R1 , ξ1 ). Then (−1)j g(ξj ) > 0,

j = 1, . . . , p.

(5.16)

Before starting the proof for Step 4, let us note that we do not lose generality by assuming f
0 in (R1 , ξ1 ); we may change (f, g) to (−f, −g) otherwise. Note also that (−1)p g(ξp ) > 0 is a contradiction to the fact that g(R2 ) = 0. Therefore this step finishes the proof of the theorem. We now prove Step 4 by an induction argument. For j = 1, suppose for contradiction that g(ξ1 )
0. Denote Ω0 = {x ∈ Ω: R1 < |x| < ξ1 }. From the equation for v0 we find that λΩ 1 (−µ + v0 + du0 ) = 0. Hence  Ω λ1 0 −µ + 2v0 + du0 + n(n + N − 2)r −2 

−2 > λΩ > 0. 1 −µ + 2v0 + du0 + n(n + N − 2)r

It follows that the strong maximum principle holds over Ω0 for the operator 

L2 := −
− µ + 2v0 + du0 + n(n + N − 2)r −2 .

It now follows from L2 g = dv0 f
0 in Ω0 and g|∂Ω0
0 that g > 0 in Ω0 ; note that by Step 3, g cannot be identically 0 in Ω0 . Similarly we find that the strong maximum principle holds over Ω0 for the operator L1 used in the proof of Step 3. It then follows from L1 f = −cu0 g < 0 in Ω0 and f |∂Ω0 = 0 that f < 0 in Ω0 . This contradiction proves (5.16) for j = 1. Suppose (5.16) holds for some i
1. If i is odd, then g(ξi ) < 0 and f  0 in (ξi , ξi+1 ). Suppose for contradiction that g(xi+1 )  0. Denote Ωi = {x ∈ Ω: ξ1 < |x| < ξi+1 }. By the same reasoning as before, the strong maximum principle is satisfied by L1 and L2 over Ωi . We have L2 g = dv0 f  0 in Ωi , and g|∂Ωi  0. Hence g < 0 in Ωi . It follows that L1 f = −cu0 g > 0 in Ωi , f |∂Ωi = 0; hence f > 0 in Ωi . This contradiction shows that we must have g(ξi+1 ) > 0. If i is even, the proof is similar. Therefore (5.16) holds for all 1  j  p. The proof is complete.  We are now ready to prove the following result. T HEOREM 5.7. Suppose that (5.12) holds. Then Σ = Σr . Moreover, Σ is a smooth curve which can be parameterized by µ ∈ (µ0 , µ0 ), and Σ contains all the radially symmetric positive solutions of (5.10). Therefore for each µ ∈ (µ0 , µ0 ), (5.10) has a unique radially symmetric positive solution. P ROOF. By the local bifurcation analysis near (µ0 , θλ , 0) and (µ0 , 0, θµ0 ), Σ and Σr coincide and consist of smooth curves near these points. Away from these two points, by

Bifurcation and related topics

205

Theorem 5.6, we can apply the implicit function theorem to conclude that Σr is a smooth curve which can be parameterized by µ. Suppose for contradiction that Σ = Σr . Since they agree near (µ0 , θλ , 0) and (µ0 , 0, θµ0 ), we can find some (µ, u, v) ∈ Σr which can be approached by a sequence (µk , uk , vk ) ∈ Σ \ Σr . But this is impossible since near (µ, u, v), due to Theorem 5.6, we can apply the implicit function theorem to conclude that all the solutions to (5.10) form a smooth curve parameterized by µ, which necessarily agrees with Σr . This proves Σ = Σr . Suppose now (µ, ˆ u, ˆ v) ˆ is an arbitrary radially symmetric positive solution of (5.10). By Theorem 5.6 we can apply the implicity function theorem in the space of radially symmetric functions to conclude that, near (µ, ˆ u, ˆ v), ˆ all the radially symmetric positive solutions of (5.10) form a smooth curve parameterized by µ. Let Σ ′ denotes a maximal such curve and define µ∗ = {µ: (µ, u, v) ∈ Σ ′ }. By (5.11), we necessarily have µ∗ ∈ [µ0 , µ0 ). If µ∗ > µ0 , then we can easily deduce that there exists some (µ∗ , u∗ , v∗ ) ∈ Σ ′ , and we can apply the implicit function theorem to continue the curve Σ ′ to µ < µ∗ , contradicting the definition of µ∗ . Hence we must have µ∗ = µ0 . We now can use the equations in (5.10) to easily show that (µ0 , θλ , 0) is an accumulating point of Σ ′ . By the local bifurcation analysis near (µ0 , θλ , 0), we see that Σ ′ agrees with Σr near this point. Now Σ ′ = Σr can  be proved in the same way as we prove Σ = Σr . R EMARK 5.8. 1. In Theorems 5.6 and 5.7, the restriction that the coefficients in (5.10) are constants is not necessary; they can be radially symmetric functions. Moreover, these results remain true if Ω is a ball. See [DLO] for details. 2. In the one-dimensional case, the method of using nodal properties of the solutions to (5.13) to prove Theorem 5.6 and hence the uniqueness of positive solutions to (5.10) seems first used by López-Gómez and Pardo [LP1]. This method can be extended to similar problems with general homogeneous boundary conditions, see [LP2]. In [Hs], a related method was used to discuss a problem with homogeneous Neumann and inhomogeneous Dirichlet boundary conditions. 3. Whether the unique radially symmetric positive solution of (5.10) is linearly stable is an open problem. Theorem 5.6 rules out the possibility of symmetry breaking bifurcation, but it does not rule out the possibility of Hopf bifurcation along Σr . In [Y], by a local bifurcation analysis, it was shown that the positive solutions are linearly stable near the semitrivial solutions. 4. If we replace the Dirichlet boundary conditions in (5.10) by Neumann boundary conditions, then it is known that the positive constant solution is a global attractor of the corresponding parabolic system with any positive initial data; this can be proved by a Lyapunov function argument and relies on the assumption that the coefficients are constants, but it works for an arbitrary Ω; see [dMR].

References [AC] S.W. Ali and C. Cosner, Models for the effects of individual size and spatial scale on competition between species in heterogeneous environments, Math. Biosci. 127 (1995), 45–76. [Am] H. Amann, Fixed point equations and nonlinear eigenvalue problems in ordered Banach spaces, SIAM Rev. 18 (1976), 620–709.

206

Y. Du

[Ar] R. Aris, The Mathematical Theory of Diffusion and Reaction in Permeable Catalysts, Oxford University Press, London (1975). [BE] J. Bebernes and D. Eberly, Mathematical Problems from Combustion Theory, Springer-Verlag, New York (1989). [BB1] J. Blat and K.J. Brown, Bifurcation of steady-state solutions in predator–prey and competition systems, Proc. Roy. Soc. Edinburgh Sect. A 97 (1984), 21–34. [BB2] J. Blat and K.J. Brown, Global bifurcation of positive solutions in some systems of elliptic equations, SIAM J. Math. Anal. 17 (1986), 1339–1353. [Bo] J.M. Bony, Principe du maximum dans les espaces de Sobolev, C. R. Math. Acad. Sci. Paris Sér. A 265 (1967), 333–336. [BIS] K.J. Brown, M.M.A. Ibrahim and R. Shivaji, S-shaped bifurcation curves, Nonlinear Anal. 5 (1981), 475–486. [CC] R.S. Cantrell and C. Cosner, Should a park be an island?, SIAM J. Appl. Math. 53 (1993), 219–252. [CCH] R.S. Cantrell, C. Cosner and V. Hutson, Ecological models, permanence and spatial heterogeneity, Rocky Mountain J. Math. 26 (1996), 1–35. [CS] A. Castro and R. Shivaji, Uniqueness of positive solutions for a class of elliptic boundary value problems, Proc. Roy. Soc. Edinburgh Sect. A 98 (1984), 267–269. [CHS] E. Conway, D. Hoff and J. Smoller, Large time behaviour of solutions of systems of nonlinear reactiondiffusion equations, SIAM J. Appl. Math. 35 (1978), 1–16. [CR1] M.G. Crandall and P.H. Rabinowitz, Bifurcation from simple eigenvalues, J. Funct. Anal. 8 (1971), 321–340. [CR2] M.G. Crandall and P.H. Rabinowitz, Bifurcation, perturbation of simple eigenvalues, and linearized stability, Arch. Ration. Mech. Anal. 52 (1973), 161–180. [Da1] E.N. Dancer, On the structure of solutions of an equation in catalysis theory when a parameter is large, J. Differential Equations 37 (1980), 404–437. [Da2] E.N. Dancer, On the indices of fixed points of mappings in cones and applications, J. Math. Anal. Appl. 91 (1983), 131–151. [Da3] E.N. Dancer, On positive solutions of some pairs of differential equations, Trans. Amer. Math. Soc. 284 (1984), 729–743. [Da4] E.N. Dancer, On positive solutions of some pairs of differential equations, II, J. Differential Equations 60 (1985), 236–258. [Da5] E.N. Dancer, The effect of domain shape on the number of positive solutions of certain nonlinear equations, J. Differential Equations 74 (1988), 120–156. [Da6] E.N. Dancer, On the existence and uniqueness of positive solutions for competing species models with diffusion, Trans. Amer. Math. Soc. 326 (1991), 829–859. [Da7] E.N. Dancer, Upper and lower stability and index theory for positive mappings and applications, Nonlinear Anal. 17 (1991), 205–217. [Da8] E.N. Dancer, Some remarks on classical problems and fine properties of Sobolev spaces, Differential Integral Equations 9 (1996), 437–446. [DD1] E.N. Dancer and Y. Du, Competing species equations with diffusion, large interactions, and jumping nonlinearities, J. Differential Equations 114 (1994), 434–475. [DD2] E.N. Dancer and Y. Du, Effects of certain degeneracies in the predator–prey model, SIAM J. Math. Anal. 34 (2002), 292–314. [DLO] E.N. Dancer, J. Lopez-Gomez and R. Ortega, On the spectrum of some linear noncooperative elliptic systems with radial symmetry, Differential Integral Equations 8 (1994), 515–523. [dP] M. del Pino, Positive solutions of a semilinear elliptic equation on a compact manifold, Nonlinear Anal. 22 (1994), 1423–1430. [De] K. Deimling, Nonlinear Functional Analysis, Springer-Verlag, Berlin (1985). [dMR] P. de Mottoni and F. Rothe, Convergence to homogeneous equilibrium state for generalized Volterra– Lotka systems, SIAM J. Appl. Math. 37 (1979), 648–663. [Du1] Y. Du, The structure of the solution set of a class of nonlinear eigenvalue problems, J. Math. Anal. Appl. 170 (1992), 567–580. [Du2] Y. Du, Exact multiplicity and S-shaped bifurcation curve for some semilinear elliptic problems from combustion theory, SIAM J. Math. Anal. 32 (2000), 707–733.

Bifurcation and related topics

207

[Du3] Y. Du, Effects of a degeneracy in the competition model, Part I and Part II, J. Differential Equations 181 (2002), 92–164. [Du4] Y. Du, Realization of prescribed patterns in the competition model, J. Differential Equations 193 (2003), 147–179. [Du5] Y. Du, Spatial patterns for population models in a heterogeneous environment, Taiwanese J. Math. 8 (2004), 155–182. [Du6] Y. Du, Asymptotic behavior and uniqueness results for boundary blow-up solutions, Differential Integral Equations 17 (2004), 819–834. [DB] Y. Du and K.J. Brown, Bifurcation and monotonicity in competition reaction–diffusion systems, Nonlinear Anal. 23 (1994), 1–13. [DG] Y. Du and Z.M. Guo, The degenerate logistic model and a singularly mixed boundary blow-up problem, Discrete Contin. Dyn. Syst., to appear. [DHs] Y. Du and S.B. Hsu, A diffusive predator–prey model in heterogeneous environment, Preprint (2003). [DH] Y. Du and Q. Huang, Blow-up solutions for a class of semilinear elliptic and parabolic equations, SIAM J. Math. Anal. 31 (1999), 1–18. [DL] Y. Du and S.J. Li, Positive solutions with prescribed patterns in some simple semilinear equations, Differential Integral Equations 15 (2002), 805–822. [DLo1] Y. Du and Y. Lou, S-shaped global bifurcation curve and Hopf bifurcation of positive solutions to a predator–prey model, J. Differential Equations 144 (1998), 390–440. [DLo2] Y. Du and Y. Lou, Proof of a conjecture for the perturbed Gelfand equation from combustion theory, J. Differential Equations 173 (2001), 213–230. [DM] Y. Du and L. Ma, Logistic type equations on RN by a squeezing method involving boundary blow-up solutions, J. London Math. Soc. 64 (2001), 107–124. [DO] Y. Du and T. Ouyang, Bifurcation from infinity induced by a degeneracy in semilinear equations, Adv. Nonlinear Stud. 2 (2002), 117–132. [FKLM] J.M. Fraile, P. Koch-Medina, J. Lopez-Gomez and S. Merino, Elliptic eigenvalue problems and unbounded continua of positive solutions of a semilinear equation. J. Differential Equations 127 (1996), 295–319. [GGLS] J. Garcia-Melian, R. Gomez-Renasco, J. Lopez-Gomez and J.C. Sabina de Lis, Pointwise growth and uniqueness of positive solutions for a class of sublinear elliptic problems where bifurcation from infinity occurs, Arch. Ration. Mech. Anal. 145 (1998), 261–289. [GNN] B. Gidas, W.-M. Ni and L. Nirenberg, Symmetry and related properties via the maximum principle, Comm. Math. Phys. 68 (1979), 209–243. [GT] D. Gilbarg and N.S. Trudinger, Elliptic Partial Differential Equation of Second Order, Springer-Verlag, Berlin (2001). [HM] S.P. Hastings and J.B. McLeod, The number of solutions to an equation from catalysis, Proc. Roy. Soc. Edinburgh Sect. A 101 (1985), 15–30. [HK] T.J. Healey and H. Kielhofer, Separation of global solution branches of elliptic systems with symmetry via nodal properties, Nonlinear Anal. 21 (1993), 665–684. [He] P. Hess, Periodic–Parabolic Boundary Value Problems and Positivity, Pitman Research Notes in Math., Vol. 247, Longman, Harlow (1991). [Hir] M. Hirsch, Stability and convergence in strongly monotone dynamical systems, J. Reine Angew. Math. 383 (1988), 1–53. [Hs] S.B. Hsu, Steady states of a system of partial differential equations modeling microbial ecology, SIAM J. Math. Anal. 14 (1983), 1130–1138. [HSW] S.B. Hsu, H.L. Smith and P. Waltman, Competitive exclusion and coexistence for competitive systems on ordered Banach spaces, Trans. Amer. Math. Soc. 348 (1996), 4083–4094. [HLM] V. Hutson, Y. Lou and K. Mischaikow, Spatial heterogeneity of resources versus Lotka–Volterra dynamics, J. Differential Equations 185 (2002), 97–136. [HLMP] V. Hutson, Y. Lou, K. Mischaikow and P. Polacik, Competing species near a degenerate limit, SIAM J. Math. Anal. 35 (2003), 453–491. [HMP] V. Hutson, K. Mischaikow and P. Polacik, The evolution of dispersal rates in a heterogeneous timeperiodic environment, J. Math. Biol. 43 (2001), 501–533.

208

Y. Du

[JL] D.D. Joseph and T.S. Lundgren, Quasilinear Dirichlet problems driven by positive sources, Arch. Ration. Mech. Anal. 49 (1973), 241–269. [Ka] Y. Kan-on, Bifurcation structure of stationary solutions of a Lotka–Volterra competition model with diffusion, SIAM J. Math. Anal. 29 (1998), 424–436. [KY] Y. Kan-on and E. Yanagida, Existence of nonconstant stable equilibria in competition diffusion equations, Hiroshima Math. J. 23 (1993), 193–221. [Kato] T. Kato, Schrödinger operators with singular potentials, Israel J. Math. 13 (1972), 133–148. [K] H. Kielhofer, Bifurcation Theory. An Introduction with Applications to PDEs, Springer-Verlag, New York (2004). [KW] K.Kishimoto and H.F. Weinberger, The spatial homogeneity of stable equilibria of some reaction– diffusion systems on convex domains, J. Differential Equations 58 (1985), 15–21. [KL] P. Koman and A.W. Leung, A general monotone scheme for elliptic systems with applications to ecological models, Proc. Roy. Soc. Edinburgh Sect. A 102 (1986), 315–325. [KLi] P. Korman and Y. Li, On the exactness of an S-shaped bifurcation curve, Proc. Amer. Math. Soc. 127 (1999), 1011–1020. [Kr] M.A. Krasnoselski, Topological Methods in the Theory of Nonlinear Integral Equations, Macmillan, New York (1965). [LN] C.S. Lin and W.-M. Ni, A counterexample to the nodal domain conjecture and a related semilinear equation, Proc. Amer. Math. Soc. 102 (1988), 271–277. [L] P.L. Lions, A remark on Bony’s maximum principle, Proc. Amer. Math. Soc. 88 (1983), 503–508. [Lop1] J. Lopez-Gomez, On the structure of the permanence region for competing species models with general diffusivities and transport effects, Discrete Contin. Dyn. Syst. 2 (1996), 525–542. [Lop2] J. Lopez-Gomez, Large solutions, metasolutions, and asymptotic behaviour of the regular positive solutions of sublinear parabolic problems, Electron. J. Differ. Equ. Conf. 5 (2000), 135–171. [LP1] J. Lopez-Gomez and R.M. Pardo, Existence and uniqueness of coexistence states for the predator–prey model with diffusion: The scalar case, Differential Integral Equations 6 (1993), 1025–1031. [LP2] J. Lopez-Gomez and R.M. Pardo, Invertibility of linear noncooperative elliptic systems, Nonlinear Anal. 31 (1998), 687–699. [LS] J. Lopez-Gomez and J. Sabina de Lis, Coexistence states and global attractivity for some convective diffusive competing species models, Trans. Amer. Math. Soc. 347 (1995), 3797–3833. [LN1] Y. Lou and W.-M. Ni, Diffusion, self-diffusion and cross diffusion, J. Differential Equations 131 (1996), 79–131. [LN2] Y. Lou and W.-M. Ni, Diffusion vs cross-diffusion: An elliptic approach, J. Differential Equations 154 (1999), 157–190. [MV] M. Marcus and L. Veron, Uniqueness and asymptotic behavior of solutions with boundary blow-up for a class of nonlinear elliptic equations, Ann. Inst. H. Poincaré Anal. Non Linéaire 14 (1997), 237–274. [Ma] H. Matano, Existence of nontrivial unstable sets for equilibriums of strongly order-preserving systems, J. Fac. Sci. Univ. Tokyo Sect. IA Math. 30 (1984), 645–673. [MM] H. Matano and M. Mimura, Pattern formation in competition–diffusion systems in nonconvex domains, Publ. Res. Inst. Math. Sci. 19 (1983), 1049–1079. [Mi] M. Mimura, Stationary patterns of some density-dependent diffusion systems with competitive dynamics, Hiroshima Math. J. 11 (1981), 621–635. [MK] M. Mimura and K. Kawasaki, Spatial segregation in competitive interaction–diffusion equations, J. Math. Biol. 9 (1980), 49–64. [MS] M. Mimura and K. Sakamoto, Multi-dimensional transition layers for an exothermic reaction–diffusion system in long cylindrical domains, J. Math. Sci. Univ. Tokyo 3 (1996), 109–179. [NS] K. Nagasaki and T. Suzuki, Spectral and related properties about the Emden–Fowler equation −
u = λeu on circular domains, Math. Ann. 299 (1994), 1–15. [N] K. Nakashima, Multiple existence of spatially inhomogeneous steady-states for competition diffusion systems, Adv. Math. Sci. Appl. 2 (1999), 973–991. [NN] W.-M. Ni and R.D. Nussbaum, Uniqueness and nonuniqueness for positive radial solutions of
u + f (u, r) = 0, Comm. Pure Appl. Math. 38 (1985), 67–108. [Ou] T. Ouyang, On the positive solutions of the semilinear equation
u + λu + hup = 0 on compact manifolds, Trans. Amer. Math. Soc. 331 (1992), 503–527.

Bifurcation and related topics

209

[OS1] T. Ouyang and J. Shi, Exact multiplicity of positive solutions for a class of semilinear problems, J. Differential Equations 146 (1998), 121–156. [OS2] T. Ouyang and J. Shi, Exact multiplicity of positive solutions for a class of semilinear problems, II, J. Differential Equations 158 (1999), 94–151. [Pao] C.V. Pao, On nonlinear reaction–diffusion systems, J. Math. Anal. Appl. 87 (1982), 165–198. [Pa] S.V. Parter, Solutions of a differential equation arising in chemical reaction processes, SIAM J. Appl. Math. 26 (1974), 687–716. [PW] M.H. Protter and H.F. Weinberger, Maximum Principles in Differential Equations, Prentice-Hall, Englewood Cliffs (1967). [Ra] P.H. Rabinowitz, Some global results for nonlinear eigenvalue problems, J. Funct. Anal. 7 (1971), 487–513. [Sa] D.H. Sattinger, Topics in Stability and Bifurcation Theory, Lecture Notes in Math., Vol. 309, SpringerVerlag, Berlin (1972). [S] J. Shi, Persistence and bifurcation of degenerate solutions, J. Funct. Anal. 169 (1999), 494–531. [Sh] R. Shivaji, Remarks on an S-shaped bifurcation curve, J. Math. Anal. Appl. 111 (1985), 374–387. [Sm] H.L. Smith, Monotone Dynamical Systems, Math. Surveys and Monographs, Vol. 41, Amer. Math. Soc. Providence, RI (1995). [T] K. Taira, Semilinear elliptic boundary-value problems in combustion theory, Proc. Roy. Soc. Edinburgh Sect. A 132 (2002), 1453–1476. [Tang] M. Tang, Exact multiplicity for semilinear elliptic Dirichlet problems involving concave and convex nonlinearities, Proc. Roy. Soc. Edinburgh Sect. A 133 (2003), 705–717. [W1] S.-H. Wang, On S-shaped bifurcation curves, Nonlinear Anal. 22 (1994), 1475–1485. [W2] S.-H. Wang, Rigorous analysis and estimates of S-shaped bifurcation curves in a combustion problem with general Arrhenius reaction-rate laws, Proc. Roy. Soc. London Ser. A 454 (1998), 1031–1048. [Wie1] H. Wiebers, S-shaped bifurcation curves of nonlinear elliptic boundary value problems, Math. Ann. 270 (1985), 555–570. [Wie2] H. Wiebers, Critical behavior of nonlinear elliptic boundary value problems suggested by exothermic reactions, Proc. Roy. Soc. Edinburgh Sect. A 102 (1986), 19–36. [Y] Y. Yamada, Stability of steady-states for prey–predator diffusion equations with homogeneous Dirichlet conditions, SIAM J. Math. Anal. 21 (1990), 327–345.

This page intentionally left blank

CHAPTER 4

Metasolutions: Malthus versus Verhulst in Population Dynamics. A Dream of Volterra Julián López-Gómez Departamento de Matemática Aplicada, Universidad Complutense de Madrid, 28040 Madrid, Spain E-mail: [email protected]

Contents 1. 2. 3. 4.

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . The main results . . . . . . . . . . . . . . . . . . . . . . . . . Some general pivotal results . . . . . . . . . . . . . . . . . . . The classical logistic equation. A priori bounds in Ω− . . . .  ⊂ Ω− . . . 4.1. The classical logistic equation: M = 0 and D  ⊂ Ω− . . . . . . . . . . . . . . . 4.2. The case M > 0 and D  ⊂ Ω− . 4.3. The radial case M = ∞ and D = BR (x0 ) with D  ⊂ Ω− . . . . . . 4.4. The case M = ∞ and D ⊂ Ω− with D 4.5. The general case M ∈ (0, ∞] and D ⊂ Ω− . . . . . . . . 5. Proofs of Theorems 2.1–2.3 . . . . . . . . . . . . . . . . . . . 6. Proof of Theorem 2.4 . . . . . . . . . . . . . . . . . . . . . . . 7. Proofs of Theorems 2.6–2.8 . . . . . . . . . . . . . . . . . . . 7.1. Finding out the boundary blow-up rate of a large solution 7.2. Two auxiliary radially symmetric problems . . . . . . . . 7.3. Proof of Propositions 7.1 and 7.2 . . . . . . . . . . . . . 7.4. Proof of Theorems 2.6 and 2.7 . . . . . . . . . . . . . . . 7.5. Proof of Theorem 2.8 . . . . . . . . . . . . . . . . . . . . 8. Relevant bibliography and further results . . . . . . . . . . . . Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

213 218 226 234 235 239 241 246 247 251 271 278 279 281 286 290 295 297 306 306

Abstract This paper analyzes how the study of the interplay within the same habitat between the most classical laws of Population Dynamics is originating the modern theory of nonlinear parabolic differential equations, where metasolutions are imperative for ascertaining the dynamics in the regimes where they cannot be described with classical solutions. HANDBOOK OF DIFFERENTIAL EQUATIONS Stationary Partial Differential Equations, volume 2 Edited by M. Chipot and P. Quittner © 2005 Elsevier B.V. All rights reserved 211

212

J. López-Gómez

Keywords: Metasolutions, Population Dynamics, Indefinite superlinear problems, Porous media MSC: 92D25, 35K57, 35J25, 35D05

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

213

1. Introduction Our attention in this monograph is focused into the problem of ascertaining the asymptotic behavior of the solutions of  ∂u   ∂t −
u = λu − a(x)f (x, u)u in Ω × (0, ∞), (1.1) u=0 on ∂Ω × (0, ∞),   in Ω, u(·, 0) = u0 > 0 where Ω is a bounded domain of RN , N
1, with smooth boundary ∂Ω, e.g., of class C 3 ,  for some µ ∈ (0, 1], satisfying the λ ∈ R, and a
0, a = 0, is a function of class C µ (Ω), following hypotheses: (Aa) The set  Ω− := x ∈ Ω: −a(x) < 0

− ⊂ Ω, whose boundary, ∂Ω− , is of class C 3 , and the is a subdomain of Ω with Ω open set − Ω0 := Ω \ Ω consists of two components, Ω0,i , i ∈ {1, 2}, such that 0,1 ∩ Ω 0,2 = ∅, Ω

0,2 ⊂ Ω, Ω

and σ1 := σ [−
, Ω0,1 ] < σ2 := σ [−
, Ω0,2 ].

(1.2)

 we denote by Throughout this paper, given a regular subdomain D of Ω and V ∈ C(D), σ [−
+ V , D] the principal eigenvalue of −
+ V in D under homogeneous Dirichlet boundary conditions. Figure 1 represents a typical situation where assumption (Aa) is fulfilled. In Figure 1 we have denoted Γ = ∂Ω,

Γ1 = ∂Ω0,1 \ Γ,

Γ2 := ∂Ω0,2 ,

∂Ω− = Γ1 ∪ Γ2 .

Thanks to Faber–Krahn inequality, (1.2) is reached if Ω0,2 has a sufficiently small Lebesgue measure (e.g., [43], Section 5 and [10], Section 10). Actually, one might think of (1.2) as a sort of hierarchical ordering size between the components Ω0,1 and Ω0,2 establishing that Ω0,1 is larger than Ω0,2 , though one should take into account that σ [−
; D] can depend on certain hidden geometrical properties of D. Setting σ0 := σ [−
, Ω],

214

J. López-Gómez

Fig. 1. Nodal configuration induced by a(x).

it is apparent, from (1.2) and the monotonicity of the principal eigenvalue with respect to the domain, that σ0 < σ1 < σ2 .

(1.3)

As for the function f (x, u), we suppose the following:  × [0, ∞)) satisfies (Af ) f ∈ C µ,1+µ (Ω  and u > 0. f (x, 0) = 0 and ∂u f (x, u) > 0 for all x ∈ Ω (Ag) There exists g ∈ C 1+µ ([0, ∞)) such that g(0) = 0,

g(u) > 0 and g ′ (u) > 0 for all u > 0,

where “ ′ ” denotes f (·, u)
g(u)

d du ,

lim g(u) = ∞,

u↑∞

and

if u
0.

Note that (Af ), (Ag) imply lim f (x, u) = ∞

u↑∞

 uniformly in x ∈ Ω.

Subsequently, we assume (Af ), (Ag), and for every constant Λ > 0 and compact set K⊂ Ω− we consider the auxiliary function hK,Λ (u) := aL,K g(u)u − Λu,

u ∈ [0, ∞),

(1.4)

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

215

where aL,K = min a > 0. K

Under assumption (Ag), it is apparent that hK,Λ has a unique positive zero. Let denote it by uK,Λ . The following assumption is crucial to get uniform a priori estimates for the solution of (1.1) within Ω− : (Ah) For every pair (K, Λ) and u > uK,Λ I (u) :=



u

∞ ' s

hK,Λ (z) dz

u

(−1/2

ds < ∞

(1.5)

and lim I (u) := 0.

(1.6)

u↑∞

Assumption (Ah) holds if g(u) = ηup−1 ,

u
0,

for some η > 0 and p > 1. Indeed, in such case, hK,Λ (u) := aL,K g(u)u − Λu = aL,K ηup − Λu,

u
0,

and hence, uK,Λ =



Λ aL,K η

1/(p−1)

.

Thus, for each s > u > uK,Λ , 

u

s

hK,Λ (z) dz =



u

s

 aL,K ηzp − Λz dz

 Λ 2  aL,K η p+1 s s − u2 = − up+1 − p+1 2

and therefore, by performing the change of variable s = uθ , we find that ∞' a

(  −1/2 Λ 2 2 I (u) = s s −u −u − ds p+1 2 u (  ∞'  Λ 2  −1/2 aL,K ηup−1 p+1 θ θ −1 = dθ −1 − p+1 2 1 

< ∞,

L,K η p+1

p+1



(1.7)

216

J. López-Gómez

since the function R(θ) defined by R(θ) :=

 Λ 2  aL,K ηup−1 p+1 −1 − θ θ −1 , p+1 2

θ
0,

satisfies R(1) = 0,

p−1

R ′ (1) = aL,K ηup−1 − Λ > aL,K ηuK,Λ − Λ = 0

and R(θ) aL,K ηup−1 > 0. = θ↑∞ θ p+1 p+1 lim

Moreover, it is easy to see that (1.7) implies (1.6). Therefore, (Ah) holds true for this choice of g. In Ecology, (1.1) models the evolution of the distribution of a single species u(x, t) randomly dispersed in the inhabiting area Ω, where λ represents the intrinsic growth rate of u and a(x) measures the crowding effects of the population in Ω− . In Ω0 , u is allowed to enjoy exponential growth according to the Malthus law. In our setting, the inhabiting area Ω is fully surrounded by completely hostile regions, because of the homogeneous Dirich u0 > 0, represents the initial let boundary conditions on ∂Ω. The function u0 ∈ C(Ω), population distribution. Consequently, (1.1) can be viewed as a sort of intermediate prototype model linking the Malthus and the Verhulst laws of population dynamics within the same inhabiting region. Indeed, if f (x, u) = u, Ω− = Ω and a(x) > 0 for each x ∈ ∂Ω, then (1.1) provides us with the classical spatial logistic equation, while it provides us with the classical spatial Malthus equation if Ω− = ∅. Our main goal in this work is ascertaining the interplay between these two angular laws of population dynamics when they arise simultaneously in a heterogeneous environment of the type illustrated in Figure 1. Although in the classical cases when Ω− ∈ {∅, Ω} the dynamics of (1.1) is governed by the nonnegative steady-states of (1.1), i.e., by the nonnegative solutions of


−
u = λu − af (·, u)u u=0

in Ω, on ∂Ω,

(1.8)

in our general setting a new class of nonclassical nondistributional generalized steadystates must be incorporated to the mathematical analysis of the problem in order to describe the asymptotic profiles of the population as time passes by. Namely, the metasolutions. Roughly spoken, the metasolutions of (1.1) are the extensions by infinity of the explosive solutions, or large solutions, of −
u = λu − af (·, u)u

(1.9)

0,1 , Ω− }. From the biological point of view, the main results of this paper in D ∈ {Ω \ Ω can be shortly summarized as follows:

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

217

• The inhabiting region Ω cannot support the species u if λ  σ0 . • The species u grows according to the Verhulst law if σ0 < λ < σ1 . 0,1 , while it is governed by the • The species u grows according to the Malthus law in Ω 0,1 if σ1  λ < σ2 . Verhulst law in Ω \ Ω • The species u grows according to the Verhulst law in Ω− , while it exhibits Malthusian − if λ
σ2 . growth in Ω \ Ω Consequently, the nature of the evolution of a single randomly distributed spatial species in a heterogeneous environment might suffer drastic changes according to the size of its intrinsic growth rate, evolving from extinction and logistic growth up to exhibit a genuine exponential growth within the most favorable regions, as in such cases most of the individuals of the population tend to abandon the most hostile areas for colonizing the favorable regions where natural resources are almost unlimited. A most detailed discussion, once stated the main mathematical results of this monograph, will be carried out in Section 2. Besides the huge intrinsic interest of analyzing (1.1), as a result of the wide variety of its applications in the applied sciences and engineering, analyzing (1.1) is imperative as well from the point of view that it is a cell model for designing more sophisticated – so, more realistic – multispecies interacting models. Actually, it has been recently shown that the asymptotic profiles of the solutions of (1.1) also provide us with the dynamics of the positive solutions of large classes of superlinear indefinite problems, where a(x) changes of sign (cf. Section 8). Therefore, the theory of metasolutions developed in this monograph should be a milestone to generate a great variety of new mathematical results in analyzing the effects of spatial heterogeneities in Chemistry, Biology, Ecology, Economy and Physics, so tremendously facilitating the understanding of the role of spatial heterogeneities in the formation and diversity of the Universe.  it is said Throughout this paper, given a subdomain D ⊂ Ω and a function u ∈ C(D),  it is said that u > v if that u > 0 if u
0 and u = 0. Accordingly, given u, v ∈ C(D),  it is said that u ≫ 0 if u(x) > 0 for each x ∈ D and u − v > 0. Also, given u ∈ C 1 (D), ∂u −1 (0), where n stands for the outward unit normal of D (x) < 0 for each x ∈ ∂D ∩ u x ∂nx  it is said that u ≫ v if u − v ≫ 0. at x ∈ ∂D. Accordingly, given u, v ∈ C 1 (D), Under our regularity assumptions, (1.1) possesses a unique smooth classical solution u(x, t) := u[λ,Ω] (x, t; u0 ) globally defined in time since −af (·, u)u  0. Moreover, by the parabolic maximum principle, u(·, t) ≫ 0 for each t > 0, since u0 > 0. The main goal of this work is ascertaining the behavior of the population distribution as time passes by, i.e., characterizing the limit L := lim u(·, t), t↑∞

(1.10)

if it exists, according to the several ranges of values where the intrinsic growth rate λ varies. In terms of L, the main results of this paper can be listed as follows:

218

J. López-Gómez

• L = 0 if λ  σ0 . • 0 < L < ∞ if σ0 < λ < σ1 . 0,1 \ ∂Ω and 0 < L < ∞ in Ω \ Ω 0,1 if σ1  λ < σ2 . • L = ∞ in Ω • L = ∞ in Ω \ Ω− and 0 < L < ∞ in Ω− if λ
σ2 . This monograph is distributed as follows. In Section 2 we state the main results and discuss their biological meaning. In Section 3 we collect the characterization of the strong maximum principle found by López-Gómez and Molina-Meyer in [53] and characterize the dynamics of a class of general parabolic problems related to (1.1). These results are crucial in the subsequent mathematical analysis. In Section 4 we give some preliminary results that will be necessary to prove the results of Section 2, among them count some classical problems of logistic type and some substantial improvements of the classical uniform a priori bounds of Keller [38] and Osserman [61]. In Sections 5–7 we will prove the results of Section 2. Finally, in Section 8 we describe the genesis and evolution of the mathematical theory of metasolutions and discuss some further, very recent, applications to porous media and indefinite superlinear problems. Throughout this paper, given two real Banach spaces X and Y and a linear continuous operator between X and Y , say L, N[L] and R[L] will stand for the null space – kernel – and the range – image – of L. 2. The main results Throughout this section we will assume that a(x) satisfies (Aa) and that f satisfies (Af ), (Ag) and (Ah), though some of the results might be valid under much weaker assumptions. Then, any weak solution of (1.9) in a smooth subdomain D ⊂ Ω must live in C 2+µ (D). The solution of (1.1) satisfies  µ  × (0, ∞) . u ∈ C 2+µ,1+ 2 Ω

Throughout this paper, given a smooth subdomain D ⊂ Ω and M ∈ [0, ∞], we consider the family of elliptic boundary value problems
−
u = λu − a(x)f (x, u)u in D, (2.1) u=M on ∂D.  if 0  M < ∞, while, in case Any weak solution u of (2.1) must satisfy u ∈ C 2+µ (D) M = ∞, a function u ∈ C 2+µ (D) is said to be a solution of (2.1) if it satisfies the differential equation in D and lim

x∈D dist(x,∂D)↓0

u(x) = ∞.

(2.2)

Throughout this paper, such solutions are called large solutions – or explosive solutions – of (1.9) in D (cf. [6,57] and the references therein). The following result characterizes the existence of large solutions of (1.9) in Ω− . These solutions will provide us with the limiting profiles of the solutions of (1.1) when λ
σ2 .

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

219

T HEOREM 2.1. Suppose M = ∞ and D = Ω− . Then, for each λ ∈ R, (2.1) possesses a max minimal and a maximal positive solution, denoted by Lmin [λ,Ω− ] and L[λ,Ω− ] , respectively – in the sense that any other positive solution L of (2.1) must satisfy max Lmin [λ,Ω− ]  L  L[λ,Ω− ] .

(2.3)

0,1 . The following result characterizes the existence of large solutions of (1.9) in Ω \ Ω These solutions will provide us with the dynamics of (1.1) for the range σ1  λ < σ2 . T HEOREM 2.2. Suppose M = ∞ and 0,1 = Ω− ∪ Ω 0,2 . D=Ω \Ω Then (2.1) possesses a positive solution if and only if λ < σ2 . Moreover, for each λ < σ2 , max (2.1) has a minimal and a maximal solution, denoted by Lmin 0,1 ] and L[λ,Ω\Ω 0,1 ] , re[λ,Ω\Ω spectively – in the sense that any other solution L of (2.1) must satisfy Lmin  [λ,Ω\Ω

0,1 ]

 L  Lmax  [λ,Ω\Ω

0,1 ]

.

(2.4)

Moreover, there exists a large solution of equation −
u = σ2 u − af (·, u)u in Ω− , say L[σ2 ,Ω− ] , such that lim Lmin  λ↑σ2 [λ,Ω\Ω0,1 ]

=



∞ L[σ2 ,Ω− ]

0,2 , in Ω in Ω− .

(2.5)

The following result characterizes the existence of positive solutions of (1.8), which equals (2.1) if M = 0 and D = Ω. These solutions will provide us with the dynamics of (1.1) within the range σ0 < λ < σ1 . T HEOREM 2.3. Problem (1.8) has a positive solution if and only if σ0 < λ < σ 1 . Moreover, it is unique, if it exists, and if we denote it by θ[λ,Ω] , then

  , lim θ[λ,Ω] = 0 in C Ω

(2.6)

λ↓σ0

and lim θ[λ,Ω] =

λ↑σ1



∞ Lmin  [σ ,Ω\Ω 1

0,1 ]

0,1 \ ∂Ω, in Ω 0,1 . in Ω \ Ω

(2.7)

220

J. López-Gómez

Furthermore, the map

 θ  , (σ0 , σ1 ) −→ C Ω

(2.8)

λ → θ (λ) := θ[λ,Ω]

is point-wise increasing and of class C 1 . The following result provides us with the dynamics of (1.1). T HEOREM 2.4. Let u(x, t) := u[λ,Ω] (x, t; u0 ) denote the unique solution of (1.1). Then  if λ  σ0 . (a) limt↑∞ u(·, t) = 0 in C(Ω)  if σ0 < λ < σ1 . (b) limt↑∞ u(·, t) = θ[λ,Ω] in C(Ω) (c) In case σ1  λ < σ2 , the following assertions are true: 0,1 \ ∂Ω. (i) limt↑∞ u(·, t) = ∞ uniformly in compact subsets of Ω  (ii) In Ω \ Ω0,1 the following estimate is satisfied Lmin  [λ,Ω\Ω

0,1 ]

 lim inf u(·, t)  lim sup u(·, t)  Lmax  [λ,Ω\Ω t↑∞

t↑∞

0,1 ]

.

(iii) If, in addition, u0 is a subsolution of (1.9) in Ω, then lim u(·, t) = Lmin  [λ,Ω\Ω

0,1 ]

t↑∞

0,1 . in Ω \ Ω

(d) In case λ
σ2 , the following assertions are true: (i) limt↑∞ u(·, t) = ∞ uniformly in compact subsets of Ω \ Ω− . (ii) In Ω− the following estimate is satisfied max Lmin [λ,Ω− ]  lim inf u(·, t)  lim sup u(·, t)  L[λ,Ω− ] . t↑∞

t↑∞

(iii) If, in addition, u0 is a subsolution of (1.9) in Ω, then lim u(·, t) = Lmin [λ,Ω− ]

t↑∞

in Ω− .

The statements of Theorem 2.4(c) and (d) can be substantially shortened by introducing the following concept. 0,1 , Ω− }. Then, a function M : Ω → D EFINITION 2.5. Suppose M = ∞ and D ∈ {Ω \ Ω [0, ∞] is said to be a metasolution of (1.9) supported in D if there exists a large solution L of (1.9) in D for which M=




∞ L

in Ω \ D, in D.

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

221

The metasolution M is said to be the minimal (resp. maximal) metasolution of (1.9) in D if L is the minimal (resp. maximal) large solution in D. The minimal and the maximal max metasolution of (1.9) in D are throughout denoted by Mmin [λ,D] and M[λ,D] , respectively. Using Definition 2.5, Theorem 2.4(c) shows that, if σ1  λ < σ2 , Mmin  [λ,Ω\Ω

0,1 ]

 lim inf u(·, t)  lim sup u(·, t)  Mmax  [λ,Ω\Ω t↑∞

t↑∞

0,1 ]

in Ω,

and actually, lim u(·, t) = Mmin  [λ,Ω\Ω

in Ω

0,1 ]

t↑∞

if, in addition, u0 is a subsolution of (1.9) in Ω. Similarly, thanks to Theorem 2.4(d), for any λ
σ2 we have that max Mmin [λ,Ω− ]  lim inf u(·, t)  lim sup u(·, t)  M[λ,Ω− ] t↑∞

in Ω.

t↑∞

Moreover, lim u(·, t) = Mmin [λ,Ω− ]

in Ω

t↑∞

if, in addition, u0 is a subsolution of (1.9) in Ω. As a result from these features, the dynamics of (1.1) is governed by the maximal classical nonnegative solution of (1.9) if λ < σ1 , by the metasolutions of (1.9) supported in 0,1 if σ1  λ < σ2 , and by the metasolutions of (1.9) supported in Ω− if λ
σ2 . Ω \Ω The following results provide us with some sufficient conditions for the uniqueness of 0,1 , Ω− }. the metasolution of (1.9) supported in D ∈ {Ω \ Ω T HEOREM 2.6. Suppose f (x, u) = up−1 ,

 × [0, ∞), (x, u) ∈ Ω

(2.9)

for some p > 1, and there exist



 β ∈ C Γ1 ; (0, ∞) and γ ∈ C Γ1 ; (0, ∞) such that

lim

x→x1 x∈Ω−

a(x) = 1 uniformly in x1 ∈ Γ1 . β(x1 )[dist(x, Γ1 )]γ (x1 )

(2.10)

Then, for each λ < σ2 , Lmin  [λ,Ω\Ω

0,1 ]

= Lmax  [λ,Ω\Ω

0,1 ]

,

(2.11)

222

J. López-Gómez

0,1 . Therefore, (1.9) possesses a unique i.e., (1.9) possesses a unique large solution in Ω \ Ω 0,1 , subsequently denoted by M[λ,Ω\Ω ] , and, due to metasolution supported in Ω \ Ω 0,1 Theorem 2.4(c), M[λ,Ω\Ω0,1 ] is a global attractor for the solutions of (1.1) if σ1  λ < σ2 . T HEOREM 2.7. Suppose (2.9) and there exist

such that

 β ∈ C ∂Ω− ; (0, ∞) lim

x→x1 x∈Ω−

 and γ ∈ C ∂Ω− ; (0, ∞)

a(x) = 1 uniformly in x1 ∈ ∂Ω− . β(x1 )[dist(x, Γ1 )]γ (x1 )

(2.12)

Then, for each λ ∈ R, max Lmin [λ,Ω− ] = L[λ,Ω− ] ,

(2.13)

i.e., (1.9) possesses a unique large solution in Ω− . Consequently, (1.9) possesses a unique metasolution supported in Ω− , subsequently denoted by M[λ,Ω− ] , and, due to Theorem 2.4(d), M[λ,Ω− ] is a global attractor for the solutions of (1.1) if λ
σ2 . The next result shows the continuity and monotonicity in λ of the metasolutions of (1.9) 0,1 , Ω− } under the assumptions of Theorem 2.7, though the result supported in D ∈ {Ω \ Ω remains valid assuming the uniqueness of the metasolution for every value of the parameter where it exists. T HEOREM 2.8. Under the assumptions of Theorem 2.7, each of the maps MΩ\Ω 

 0,1 (−∞, σ2 ) −→ C Ω; (0, ∞] ,

(2.14)

λ → MΩ\Ω0,1 (λ) := M[λ,Ω\Ω0,1 ]

and MΩ

 − R −→ C Ω; (0, ∞] ,

(2.15)

λ → MΩ− (λ) := M[λ,Ω− ]

is continuous and point-wise increasing. Moreover, lim θ (λ) = MΩ\Ω0,1 (σ1 )

λ↑σ1

and

where θ is the solution map defined in (2.8).

lim MΩ\Ω0,1 (λ) = MΩ− (σ2 ),

λ↑σ2

(2.16)

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

223

Fig. 2. The dynamics of (1.1). Classical solutions and metasolutions.

In Figure 2 we have represented the dynamics of (1.1) under the assumptions of Theorem 2.7. Most precisely, we have drown the diagram of classical solutions and significant metasolutions of (1.9) together with their respective attracting properties. We are plotting the parameter λ versus the value of the generalized solution, u(x), at some point x ∈ Ω− , where all these generalized solutions are finite, so that the diagram cannot exhibit any bifurcation from infinity. The diagram shows four different kind of solutions. The λ-axis represents u = 0, which, according to Theorem 2.4(a), is a global attractor if λ  σ0 , while it is linearly unstable for any λ > σ0 . Then, we have represented the positive solution θ[λ,Ω] , which bifurcates from u = 0 at λ = σ0 and it is point-wise increasing for λ ∈ (σ0 , σ1 ) until it reaches the metasolution M[σ1 ,Ω\Ω0,1 ] at λ = σ1 . Due to Theorem 2.4(b), θ[λ,Ω] is a global attractor of (1.1) for each λ ∈ (σ0 , σ1 ). Then, we have represented the curve λ → M[λ,Ω\Ω0,1 ] (x), which, according to Theorem 2.8, is continuous and increasing in its definition interval (−∞, σ2 ). Thanks to Theorem 2.4(c), M[λ,Ω\Ω0,1 ] is a global attractor for the solutions of (1.1) if λ ∈ [σ1 , σ2 ), and, due to Theorem 2.2, it approximates the metasolution M[σ2 ,Ω− ] as λ ↑ σ2 . As for any λ < σ1 the dynamics of (1.1) is described by the classical nonnegative steady states of (1.1), in such range the metasolution M[λ,Ω\Ω0,1 ] must be unstable from below. Consequently, the value λ = σ1 provides us with the critical value of the parameter 0,1 changes. Finally, where the attractive character of the metasolution supported in Ω \ Ω Figure 2 shows the curve λ → M[λ,Ω− ] (x), which, according to Theorem 2.8, is continuous and increasing in R. Thanks to Theorem 2.4, M[λ,Ω− ] is a global attractor for the solutions of (1.1) if λ
σ2 , while it is unstable from below if λ < σ2 . So, λ = σ2 is the critical value of the parameter where the attractive character of M[λ,Ω− ] changes. Figure 3 shows the corresponding asymptotic profiles of the

224

J. López-Gómez

Fig. 3. The limiting profiles of the solutions of (1.1).

solutions of (1.1) according to the range of variation of the parameter λ. Although for a general nonlinearity f (x, u) satisfying (Af ), (Ag) and (Ah) our theory does not guarantees the uniqueness of the metasolution, due to Theorem 2.4, Figure 2 still provides us with the dynamics of (1.1) when u0 is a subsolution of (1.9), though in this case one should think of the minimal metasolution curves, which might exhibit a number of jumps as a result of the eventual nonuniqueness of the metasolutions. Nevertheless, though we could not prove 0,1 and Ω− must be unique it yet, we conjecture that the metasolutions supported in Ω \ Ω for a general a(x) satisfying (Aa) and f (x, u), not necessarily of the special form (2.9). Theorem 2.4 is of fundamental interest from the point of view of the applications of the abstract mathematical theory developed here to population dynamics as it provides us with the simultaneous effects of incorporating both Malthus and Verhulst laws within the same natural environment, which should be extremely realistic from the point of view of applications to real world models. Rather naturally, the density of the species should be severely limited in the regions where natural resources are drastically limited, though it might be certainly unlimited within the regions where natural resources are sufficiently abundant to maintain a huge population. A mechanism explaining why agriculture facilitated the emergence of human groups whose size gradually increased during the last 10 thousands years – nothing at the human evolution scale – until originating the extremely densely populated areas that we inhabit today. Simultaneously, in unfavorable areas, where agriculture was

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

225

not possible, or simply extremely difficult, like in the rain forest areas, the level of human population has been controlled almost at the original levels. As a consequence from Theorem 2.4, if the intrinsic growth rate of the species, measured by λ, is below the threshold σ0 , then the inhabiting area cannot support the species u, which is driven to extinction. When λ ∈ (σ0 , σ1 ), then the inhabiting area is able to maintain the species u around the critical level θ[λ,Ω] , independently of the size of the initial population u0 . So, Ω cannot maintain an arbitrarily large population. Therefore, (1.1) exhibits a genuine logistic behavior if λ < σ1 . Quite surprisingly from the mathematical point of view – for a reason to be explained in Section 8 – in the interval λ ∈ [σ1 , σ2 ) the population must be limited in the region 0,1 , while, as an effect from dispersion to the less hostile region Ω0,1 , the population Ω \Ω 0,1 can grow arbitrarily within Ω0,1 . Thus, (1.1) exhibits a sort of logistic growth in Ω \ Ω and a genuine exponential growth in Ω0,1 if λ ∈ [σ1 , σ2 ). Indeed, thanks to the parabolic maximum principle, u1 := u[λ,Ω] (·, 1; u0 ) ≫ 0 and, for each t
0, u[λ,Ω] (·, t + 1; u0 ) = u[λ,Ω] (·, t; u1 ) > u[λ,Ω0,1 ] (·, t; u1 ),

(2.17)

where u[λ,Ω0,1 ] (x, t; u1 ) stands for the solution of the linear problem  ∂u   ∂t −
u = λu u=0   u(·, 0) = u1

in Ω0,1 × (0, ∞),

on ∂Ω0,1 × (0, ∞), in Ω0,1 ,

(2.18)

which is given by

u[λ,Ω0,1 ] (x, t; u1 ) = et (λ+
) u1 .

(2.19)

Now, suppose λ > σ1 and denote by ϕ0,1 the principal eigenfunction associated with σ [−
, Ω0,1 ]. Then, since u1 ≫ 0, there exists α > 0 such that u1 > αϕ0,1 and hence, it follows from (2.17) and (2.19) that u[λ,Ω] (·, t + 1; u0 ) > αet (λ+
) ϕ0,1 = αe(λ−σ1 )t ϕ0,1 , which shows the exponential growth of the population within Ω0,1 . Similarly, when λ > σ2 the population must be controlled within Ω− , as a result of the limitation of the natural resources available there in, though the individual of the species can disperse to the most favorable areas Ω0,1 and Ω0,2 , where the population density can

226

J. López-Gómez

be arbitrarily large; actually growing at the respective rates et (λ−σ1 ) and et (λ−σ2 ) . Consequently, the principal eigenvalues σ1 and σ2 might be though as sort of measuring parameters of the quality of each of the refuge patches of the environment. 3. Some general pivotal results The following characterization of the strong maximum principle goes back to [53] and [43] (cf. [48] for an extremely sharp version of the theorem). T HEOREM 3.1. Suppose D is an open subdomain of RN , N
1, with smooth boundary  Then, the following assertions are equivalent. and V ∈ C µ (D). (a) σ [−
+ V , D] > 0.  such that h > 0 in D, (b) There exists a function h ∈ C 2 (D) ∩ C(D) (−
+ V )h
0 in D, and either h|∂D > 0, or (−
+ V )h > 0 in D – such a function h is called a positive strict supersolution of −
+ V in D under Dirichlet boundary conditions. (c) The operator −
+ V satisfies the strong maximum principle in D, i.e., for every  g ∈ C 2+µ (∂D), such that f
0, g
0, (f, g) = (0, 0), and any u ∈ C 2+µ (D)  f ∈ C µ (D), satisfying
(−
+ V )u = f in D, u=g on ∂D, one has that u ≫ 0 in D.  × [0, ∞)) and Throughout the remaining part of this section we suppose f ∈ C µ,µ (Ω consider a smooth subdomain D of Ω such that D ∩ Ω− = ∅.

(3.1)

As an immediate consequence from the abstract theory developed by Amann in [1] and [2], the following result holds. T HEOREM 3.2. Suppose g ∈ C 2+µ (∂D) and
−
u = λu − af (·, u)u in D, u=g on ∂D

(3.2)

 and a supersolution u¯ ∈ C 2+µ (D)  such that u  u. possesses a subsolution u ∈ C 2+µ (D) ¯ 2+µ  Then (3.2) possesses a solution u ∈ C (D) such that ¯ u  u  u. ¯ Moreover, (3.2) has a minimal and a maximal solution in [ u, u].

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

227

Note that if u (resp. u) ¯ is a strict subsolution (resp. supersolution) of (3.2), then any ¯ must satisfy u < u  u¯ (resp. u  u < u). ¯ solution u of the problem in [ u, u] As an easy consequence from Theorem 3.2, the following result holds. T HEOREM 3.3. Suppose u, u¯ ∈ C 2+µ (D), satisfy −
u  λu − af (·, u )u lim

dist(x,∂D)↓0

u(x) = ∞

−
u¯
λu¯ − af (·, u) ¯ u¯

and

and

lim

dist(x,∂D)↓0

in D,

u(x) ¯ = ∞,

and u  u¯ in D. Then, the singular boundary value problem,


−
u = λu − af (·, u)u u=∞

in D, on ∂D,

(3.3)

¯ possesses a solution u ∈ C 2+µ (D) such that u  u  u. P ROOF. For each sufficiently large n
1, say n
n0 , we consider


1 Dn := x ∈ D: dist(x, ∂D) > . n The integer n0
1 must be chosen so that ∂Dn inherits the regularity of ∂D. Thanks to Theorem 3.2, for each n
n0 , the problem


−
u = λu − af (·, u)u

u=

u+u¯ 2

in Dn , on ∂Dn

n ) such that possesses a solution un ∈ C 2+µ (D ¯ Dn u|Dn  un  u|

in Dn .

Thanks to these estimates, there exists a subsequence {unm }m
1 of {un }n
n0 such that lim unm − u0 C 2+µ (D n ) = 0

m→∞

0

n0 ) of for some solution u0 ∈ C 2+µ (D 

−
u = λu − af (·, u)u

u=

u+u¯ 2

in Dn0 , on ∂Dn0 .

228

J. López-Gómez

Now, consider the new sequence {unm |Dn0 +1 }m
1 . The previous argument also shows the existence of a subsequence of {unm |Dn0 +1 }m
1 , labeled again by nm , such that, for some n0 +1 ), u1 ∈ C 2+µ (D lim unm − u1 C 2+µ (D n

m→∞

0 +1

)

= 0.

Necessarily, u1 |Dn0 = u0 . Repeating this procedure infinitely many times, the point-wise limit of the diagonal sequence provides us with a solution of (3.3) satisfying all requirements.  The following result collects some important properties that are going to be used throughout the remaining of this work.  u¯ > 0, L EMMA 3.4. Suppose f satisfies (Af ), g ∈ C 2+µ (∂D), g
0, and u¯ ∈ C 2+µ (D), ∂ u¯ is a supersolution of (3.2). Then, u¯ ≫ 0, i.e., u(x) ¯ > 0 for each x ∈ D and ∂nx (x) < 0 for each x ∈ u¯ −1 (0) ∩ ∂D, where nx stands for the outward unit normal at x ∈ ∂D. In particular, any positive solution u of (3.2) satisfies u ≫ 0. Moreover, % & σ −
− λ + af (·, u), ¯ D
0.

Actually, if g = 0 and u¯ is a solution of (3.2), then % & λ = σ −
+ af (·, u), ¯ D .

(3.4)

Furthermore, for each κ > 1, κ u¯ also provides us with a supersolution of (3.2). ¯ ∂D
0 and P ROOF. Since u|

 −
− λ + af (·, u) ¯ u¯
0 in D,

u¯ > 0 provides us with a positive supersolution of −
− λ + af (·, u) ¯ in D under Dirichlet boundary conditions, and two different situations can occur. If either g > 0 on ∂D, or g = 0 on ∂D but u¯ is a positive strict supersolution of −
− λ + af (·, u) ¯ in D under Dirichlet boundary conditions, then it follows from Theorem 3.1 that % & σ −
− λ + af (·, u), ¯ D >0

(3.5)

and u¯ ≫ 0. If g = 0 on ∂D and u¯ is not a strict supersolution of −
− λ + af (·, u) ¯ in D, then u¯ provides us with a positive eigenfunction associated with the principal eigenvalue % & σ −
− λ + af (·, u), ¯ D =0

(3.6)

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

229

and hence, by Krein–Rutman theorem, u ≫ 0. Note that in such case u¯ is a solution of (3.2) and that (3.4) holds from (3.6). Now, pick κ > 1. Then, κ u| ¯ ∂D
u| ¯ ∂D
g, and, in D, we have that −
(κ u) ¯
λκ u¯ − af (·, u)κ ¯ u¯
λκ u¯ − af (·, κ u)κ ¯ u, ¯ since, by (Af ), f (·, κ u) ¯ > f (·, u), ¯ 

which concludes the proof.

The following theorem will extraordinarily simplify the mathematical analysis of the next sections. T HEOREM 3.5. Suppose f satisfies (Af ), g ∈ C 2+µ (∂D), g
0, λ > σ [−
, D] if g = 0, and (3.2) possesses a supersolution u¯ > 0. Then, (3.2) has a unique positive solution. Moreover, if we denote it by θ[λ,D,g] , then for any positive subsolution (resp. supersolution) u (resp. u) ¯ of (3.2), one has that u  θ[λ,D,g] (resp. θ[λ,D,g]  u). ¯ Furthermore, for each u0 > 0,   (3.7) lim u[λ,D,g] (·, t; u0 ) − θ[λ,D,g] C (D)  = 0, t↑∞

where u[λ,D,g] (x, t; u0 ) stands for the unique solution of the parabolic problem  ∂u   ∂t −
u = λu − af (·, u)u in D × (0, ∞), u=g   u(·, 0) = u0

on ∂D × (0, ∞), in D.

(3.8)

Suppose (Af ), (3.2) has a supersolution u¯ > 0, g = 0 and λ  σ [−
, D]. Then, (3.2) cannot admit a positive subsolution and   (3.9) lim u[λ,D,g] (·, t; u0 )C (D)  = 0. t↑∞

P ROOF. In case g > 0, u := 0 provides us with a strict subsolution of (3.2), and hence, (0, u) ¯ provides us with an ordered sub–supersolution pair. Thus, thanks to Theorem 3.2, (3.2) possesses a solution 0 < u  u. Due to Lemma 3.4, u ≫ 0. Suppose g = 0 and λ > σ [−
, D]. Let ϕ ≫ 0 denote any principal eigenfunction associated with σ [−
, D]. Then, for each sufficiently small ε > 0, the function u := εϕ provides us with a positive strict subsolution of (3.2). Indeed, εϕ|∂D = 0 and, in D, −
(εϕ) = εσ [−
, D]ϕ < λεϕ − af (·, εϕ)εϕ

230

J. López-Gómez

for sufficiently small ε > 0, since σ [−
, D] < λ and   lim af (·, εϕ)C (D)  = 0. ε↓0

Fix one of these values of ε. Due to Lemma 3.4, u ≫ 0 and κ u¯ is a supersolution of (3.2) for every κ > 1. Pick a sufficiently large κ > 1 so that εϕ < κ u. ¯

Then, (εϕ, κ u) ¯ provides us with an ordered sub–supersolution pair of (3.2) and hence, thanks to Theorem 3.2, (3.2) possesses a solution u such that εϕ < u  κ u. ¯ Note that, due to Lemma 3.4, u ≫ 0. To show the uniqueness of the positive solution of (3.2) we proceed by contradiction. Suppose (3.2) has two different positive solutions u1 = u2 . The previous analysis shows that there exist ε > 0, κ > 1, and a strict subsolution u ∈ {0, εϕ} such that u < min{u1 , u2 } < max{u1 , u2 }  κ u. ¯ ¯ Let u∗ and u∗ denote the minimal and the maximal positive solutions of (3.2) in [ u, κ u]. Necessarily, u < u∗  min{u1 , u2 } < max{u1 , u2 }  u∗  κ u¯ and therefore, (3.2) possesses two ordered positive solutions, u∗ < u∗ . Setting w := u∗ − u∗ > 0, the following linear boundary value problem is satisfied


(−
− λ + V )w = 0 in D, w=0 on ∂D,

(3.10)

where V is the potential defined by 

1

  ∂f ·, tu∗ + (1 − t)u∗ tu∗ + (1 − t)u∗ dt 0 ∂u  1

 +a f ·, tu∗ + (1 − t)u∗ dt.

V := a

0

(3.11)

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

231

By Lemma 3.4, u∗ ≫ 0 and u∗ ≫ 0. Hence, it follows from (Af ) and (3.1) that V >a



0

1

 f ·, tu∗ + (1 − t)u∗ dt
af (·, u∗ ),

since u∗ > u∗ . In particular, by the monotonicity of the principal eigenvalue with respect to the potential, we find that % & σ [−
− λ + V , D] > σ −
− λ + af (·, u∗ ), D .

(3.12)

Thus, thanks to (3.4), (3.12) gives

σ [−
− λ + V , D] > 0. As the principal eigenvalue is dominant, we conclude from (3.10) that w = 0, which contradicts w > 0. This contradiction concludes the proof of the uniqueness of the positive solution. Subsequently, we denote by θ[λ,D,g] the unique positive solution of (3.2). Suppose u > 0 is a subsolution of (3.2). Then, since u¯ ≫ 0, for sufficiently large κ > 1, we have that u < κ u¯ and hence, by Theorem 3.2, (3.2) has a positive solution in [ u, κ u]. ¯ By the uniqueness, u  θ[λ,D,g]  κ u¯ and, in particular, u  θ[λ,D,g] . Suppose u¯ > 0 is a supersolution of (3.2). By Lemma 3.4, u¯ ≫ 0. If g > 0, then, u = 0 provides us with a strict subsolution of (3.2) and, due to Theorem 3.2, (3.2) has a positive solution u in [0, u]. ¯ By the uniqueness, θ[λ,D,g]  u. ¯

(3.13)

Similarly, if g = 0, then, for sufficiently small ε > 0, u := εϕ provides us with a strict subsolution of (3.2). Thus, if ε is chosen so that εϕ < u, ¯ by Theorem 3.2, (3.2) has a positive solution u in [εϕ, u]. ¯ Therefore, by the uniqueness, (3.13) as well holds. Now, we will prove (3.7). First, we suppose that g = 0 and λ > σ [−
, D]. Since u0 > 0, by the parabolic maximum principle, for each t > 0 we have that u[λ,D,g] (·, t; u0 ) ≫ 0. Now, pick a sufficiently small ε > 0 and a sufficiently large κ > 1 so that (εϕ, κ u) ¯ be an ordered sub–supersolution pair of (3.2) for which εϕ < u[λ,D,g] (·, 1; u0 ) < κ u. ¯

232

J. López-Gómez

Note that κ > 1 can be chosen so that κ u¯ > θ[λ,D,g] , and hence, by the uniqueness of the positive solution, κ u¯ must be a positive strict supersolution. Now, thanks again to the parabolic maximum principle, for each t > 0 we have that

 u[λ,D,g] (·, t; εϕ)  u[λ,D,g] ·, t; u[λ,D,g] (·, 1; u0 ) = u[λ,D,g] (·, t + 1; u0 )  u[λ,D,g] (·, t; κ u). ¯

(3.14)

Moreover, since εϕ is a strict subsolution, as t grows, u[λ,D,g] (·, t; εϕ) increases approximating the minimal positive solution of (3.2) in [εϕ, κ u], ¯ while, since κ u¯ is a strict super¯ decreases approximating the maximal positive solution of (3.2) solution, u[λ,D,g] (·, t; κ u) in [εϕ, κ u] ¯ (cf. [65]). As θ[λ,D,g] is the unique positive solution of (3.2), passing to the limit as t ↑ ∞ in (3.14) gives lim u[λ,D,g] (·, t; u0 ) = θ[λ,D,g]

t↑∞

  , in C D

so concluding the proof of the theorem in this case. Now, suppose g > 0 and pick a sufficiently large κ > 1 for which κ u¯ > θ[λ,D,g]

and 0 < u[λ,D,g] (·, 1; u0 ) < κ u. ¯

Then, arguing as above, we find that

 u[λ,D,g] (·, t; 0)  u[λ,D,g] ·, t; u[λ,D,g] (·, 1; u0 ) = u[λ,D,g] (·, t + 1; u0 )

 u[λ,D,g] (·, t; κ u), ¯

(3.15)

and similarly, (3.7) follows by passing to the limit as t ↑ ∞ in (3.15). Finally, suppose g = 0 and λ  σ [−
, D].

(3.16)

We claim that (3.2) cannot admit a positive subsolution; in particular, it cannot admit a positive solution. To prove this feature we proceed by contradiction. So, suppose (3.2) possesses a positive subsolution u > 0. Since u¯ ≫ 0, there exists κ > 1 such that u < κ u¯ and therefore, due to Theorem 3.2, (3.2) possesses a positive solution u. Moreover, thanks to Lemma 3.4, we have u ≫ 0 and % & λ = σ −
+ af (·, u), D .

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

233

Since u ≫ 0, it follows from (Af ) and (3.1) that af (·, u) > 0 in D and hence, thanks to the monotonicity of the principal eigenvalue with respect to the potential, % & λ = σ −
+ af (·, u), D > σ [−
, D],

which contradicts (3.16). Therefore, (3.2) cannot admit a positive subsolution in such case. In order to show that, under (3.16), condition (3.9) is satisfied, we consider κ > 1 such that 0 < u[λ,D,g] (·, 1; u0 ) < κ u. ¯ Then, for each t > 0, we have that 0 < u[λ,D,g] (·, t + 1; u0 ) < u[λ,D,g] (·, t; κ u). ¯

(3.17)

As κ u¯ is a supersolution of (3.2), u[λ,D,g] (·, t; κ u) ¯ decreases approximating as t ↑ ∞ the maximal nonnegative solution of (3.2) in [0, κ u], ¯ which is the zero solution. Therefore, passing to the limit as t ↑ ∞ in (3.17) shows (3.9) and completes the proof.  Actually, the following strong comparison result is satisfied. L EMMA 3.6. Suppose f satisfies (Af ), g ∈ C 2+µ (∂D), g
0, λ > σ [−
, D] if g = 0, and (3.2) possesses a supersolution u¯ > 0. Then, for any positive strict subsolution (resp. ¯ of (3.2), one has that u ≪ θ[λ,D,g] (resp. θ[λ,D,g] ≪ u), ¯ where supersolution) u (resp. u) θ[λ,D,g] is the unique positive solution of (3.2). P ROOF. Suppose u > 0 is a strict subsolution of (3.2). Then, due to Theorem 3.5, u  θ[λ,D,g] , and therefore, u < θ[λ,D,g] , since u cannot be a solution. Consequently, w := θ[λ,D,g] − u > 0.

(3.18)

Now, adapting the uniqueness argument of the proof of Theorem 3.5, it readily follows that


(−
− λ + V )w = 0 in D, w
0 on ∂D,

where V is the potential defined by 

1

  ∂f ·, tθ[λ,D,g] + (1 − t)u tθ[λ,D,g] + (1 − t)u dt 0 ∂u  1 

+a f ·, tθ[λ,D,g] + (1 − t)u dt.

V := a

0

(3.19)

234

J. López-Gómez

Now, we have to distinguish between two different situations. If w|∂D > 0, then w > 0 provides us with a positive strict supersolution of −
− λ + V in D under homogeneous Dirichlet boundary conditions. Thus, due to Theorem 3.1, w ≫ 0 in D, and therefore u ≪ θ[λ,D,g] . Now, suppose w|∂D = 0. Then, w > 0 is a principal eigenfunction associated with σ [−
− λ + V , D] = 0 and, due to Krein–Rutman theorem, w ≫ 0, which concludes the proof.



4. The classical logistic equation. A priori bounds in Ω− In this section we analyze the dynamics of  ∂u   ∂t −
u = λu − a(x)f (x, u)u u=M   u(·, 0) = u0 > 0

in D × (0, ∞),

on ∂D × (0, ∞), in D,

(4.1)

where M ∈ [0, ∞) is a constant, D is a smooth subdomain of Ω such that D ⊂ Ω− ,

(4.2)

and f satisfies (Af ) and (Ag). The main result establishes that the dynamics of (4.1) is governed by its maximal nonnegative steady-state, which is the maximal nonnegative solution of
−
u = λu − a(x)f (x, u)u in D, (4.3) u=M on ∂D. Actually, for each M > 0, (4.3) possesses a unique positive solution which is a global attractor for the solutions of (4.1). Throughout the rest of this paper it will be denoted by θ[λ,D,M] . It turns out that this solution satisfies lim θ[λ,D,M] = 0 if λ  σ [−
, D],

M↓0

while θ[λ,D,0] := lim θ[λ,D,M] M↓0

if λ > σ [−
, D],

provides us with the unique positive solution of (4.3) for M = 0. So, there is a continuous transition between both dynamics as M perturbs from zero.  ⊂ Ω− is much easier than the analysis of the general The analysis of (4.1) in case D case when (4.2) is satisfied, where a might partially, or totally, vanish on ∂D, as a result of

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

235

 ⊂ Ω− , the fact that sufficiently large positive constants provide us with supersolutions if D while no positive constant can be a supersolution if a vanishes on ∂D and λ > 0. Thus,  ⊂ Ω− . in the first two sections we will focus our attention into the simplest case when D Then, in Section 4.3, we will assume that, in addition, f satisfies (Ah) in order to refine the classical a priori bounds of Keller [38] and Osserman [61]. These bounds will allow us to obtain the existence of a minimal and a maximal positive solution for the singular problem
−
u = λu − a(x)f (x, u)u in D, (4.4) u=∞ on ∂D. As a consequence from these bounds, it will be apparent that θ[λ,D,∞] := lim θ[λ,D,M] M↑∞

provides us with the minimal positive solution of (4.4). Finally, in Section 4.4, these results will be used to study the general case when condition (4.2) is satisfied. The following consequence from Lemma 3.4 will be very useful.  u > 0, is a solution of (4.3) for some M ∈ [0, ∞). L EMMA 4.1. Suppose u ∈ C 2+µ (D), Then u ≫ 0. Moreover, % & λ = σ −
+ af (·, u), D (4.5)

if M = 0. In case M = ∞, any solution u ∈ C 2+µ (D), u > 0, must satisfy u(x) > 0 for all x ∈ D.

P ROOF. The fact that u ≫ 0 for any positive solution of (4.3) with M ∈ [0, ∞) is a consequence from Lemma 3.4, and (4.5) is a consequence from (3.4). Now, suppose M = ∞. Then, u provides us with a positive supersolution of −
− λ + af (·, u) in any subdomain of the form Dn := {x ∈ D: dist(x, ∂D) > 1/n}, for sufficiently large n, and hence, thanks again to Lemma 3.4, u ≫ 0 in Dn , which concludes the proof.   ⊂ Ω− 4.1. The classical logistic equation: M = 0 and D The following result provides us with the existence and the uniqueness of the positive solution to (4.3) in this case. Note that aL,D := min a > 0.  D

(4.6)

 ⊂ Ω− , M = 0, and f satisfies (Af ) and (Ag). Then, (4.3) T HEOREM 4.2. Suppose D possesses a positive solution if and only if λ > σ [−
, D]. Moreover, it is unique, and strongly positive, if it exists. Furthermore, if we denote it by θ[λ,D,0] and u[λ,D,0] (x, t; u0 ) stands for the unique solution of (4.1), then

236

J. López-Gómez

 if λ  σ [−
, D]; (a) limt↑∞ u[λ,D,0] (·, t; u0 ) = 0 in C(D)  if λ > σ [−
, D]. (b) limt↑∞ u[λ,D,0] (·, t; u0 ) = θ[λ,D,0] in C(D) P ROOF. It is a direct consequence from Theorem 3.5, since sufficiently large positive constants provide us with positive supersolutions of (4.3), by (Ag).  The following result establishes that the set of positive solutions of (4.3) consists of a differentiable curve emanating from u = 0 at the value of the parameter λ = σ [−
, D], where the attractive character of the steady state u = 0, as a solution of (4.1), is lost.  ⊂ Ω− , M = 0, and f satisfies (Af ) and (Ag). Then the P ROPOSITION 4.3. Suppose D solution map

 θ

  , σ [−
, D], ∞ −→ C D λ → θ (λ) := θ[λ,D,0]

(4.7)

is of class C 1 and point-wise increasing. Actually, θ (λ) ≫ θ (µ)

if λ > µ > σ [−
, D].

Moreover, θ (λ) bifurcates from (λ, u) = (λ, 0) at λ = σ [−
, D], i.e., lim

λ↓σ [−
,D]

θ (λ) = 0.

(4.8)

P ROOF. The solutions of (4.3) are the zeros of the nonlinear operator



  → C0 D  F : R × C0 D

defined by

% & F(λ, u) := u − (−
)−1 λu − af (·, u)u ,

where (−
)−1 stands for the resolvent operator of −
in D under homogeneous Dirichlet boundary conditions. The operator F is of class C 1 and, by elliptic regularity, F(λ, ·) is a nonlinear compact perturbation of the identity for each λ ∈ R. Moreover, F(λ, 0) = 0,

λ ∈ R,

 and, for each (λ, u) ∈ R × C0 (D), Du F(λ, 0)u = u − (−
)−1 (λu). Thus, Du F(λ, 0) is a Fredholm analytic pencil of index zero whose spectrum consists of the eigenvalues of −
. In particular, %

& N Du F σ [−
, D], 0 = span[ϕ],

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

237

where ϕ ≫ 0 is any principal eigenfunction of σ [−
, D]. We claim that

 %

& Dλ Du F σ [−
, D], 0 ϕ ∈ / R Du F σ [−
, D], 0 .

(4.9)

Consequently, the transversality condition of Crandall and Rabinowitz [19] holds true. To prove (4.9) we proceed by contradiction assuming that

 %

& Dλ Du F σ [−
, D], 0 ϕ = −(−
)−1 ϕ ∈ R Du F σ [−
, D], 0 .

 such that Then there exists u ∈ C0 (D)

 u − (−
)−1 σ [−
, D]u = −(−
)−1 ϕ. 2+µ

By elliptic regularity, u ∈ C0

 and (D)

 −
− σ [−
, D] u = −ϕ.

Multiplyingthis equation by ϕ, integrating in D and applying the formula of integration by parts gives D ϕ 2 = 0, which is impossible. This contradiction shows the validity of (4.9) and therefore, by the main theorem of [19], (λ, u) = (σ [−
, D], 0) is a bifurcation point from (λ, u) = (λ, 0) to a smooth curve of positive solutions of (4.3). By the uniqueness of the positive solution, as a result from Theorem 4.2, (4.8) holds true. Subsequently we suppose that (λ, u) = (λ0 , u0 ) is a positive solution of (4.3). Then F(λ0 , u0 ) = 0 and, by Lemma 4.1, u0 ≫ 0 and % & λ0 = σ −
+ af (·, u0 ), D .

(4.10)

 Differentiating with respect to u we have that, for each u ∈ C0 (D), ' ( ∂f Du F(λ0 , u0 )u = u − (−
)−1 λ0 u − a (·, u0 )u0 u − af (·, u0 )u . ∂u In particular, Du F(λ0 , u0 ) is a Fredholm operator of index zero. Moreover, it is a linear  for which topological isomorphism, since it is inyective. Indeed, if there exists u ∈ C0 (D) −1

u − (−
)

'

( ∂f λ0 u − a (·, u0 )u0 u − af (·, u0 )u = 0, ∂u 2+µ

then, by elliptic regularity, u ∈ C0 

 and (D)

∂f −
− λ0 + a (·, u0 )u0 + af (·, u0 ) u = 0 ∂u

in D.

(4.11)

238

J. López-Gómez

On the other hand, thanks to (4.10) and (Af ), we have that ( ' ∂f σ −
− λ0 + a (·, u0 )u0 + af (·, u0 ), D ∂u & % > σ −
− λ0 + af (·, u0 ), D = 0.

Consequently, it follows from (4.11) that u = 0. Therefore, Du F(λ0 , u0 ) is a linear topological isomorphism, and the uniqueness of the positive solution, as a consequence from Theorem 4.2, combined with the implicit function theorem shows the regularity of the map θ defined by (4.7). Finally, differentiating the identity

 F λ, θ(λ) = 0,

λ > σ [−
, D],

with respect to λ gives

−1

Dλ θ = (−
)

 ∂f θ + λDλ θ − a (·, θ)θDλ θ − af (·, θ)Dλ θ ∂u

or equivalently, 

−
− λ + a

∂f (·, θ)θ + af (·, θ) Dλ θ = θ. ∂u

As θ ≫ 0 and for each λ > σ [−
, D], ' ( ∂f σ −
− λ + a (·, θ)θ + af (·, θ), D > 0, ∂u it follows from Theorem 3.1 that  

 −1 ∂f Dλ θ (λ) = −
− λ + a ·, θ(λ) θ (λ) + af ·, θ(λ) θ (λ) ≫ 0, ∂u which concludes the proof.



Note that if λ > µ > σ [−
, D], then θ[µ,D,0] is a strict subsolution of (4.3) and therefore, the relation θ[µ,D,0] ≪ θ[λ,D,0] also follows from Lemma 3.6. In Figure 4 we have illustrated the results from Theorem 4.2 and Proposition 4.3. For a given value x ∈ D, we have represented the curve λ → θ[λ,D,0] (x). It bifurcates from the horizontal axis at the value of the parameter λ = σ [−
, D] and it increases for all further values of λ. The direction of the arrows indicate the flow of (4.1). According to Theorem 4.2, as time grows to infinity u[λ,D,0] (x, t; u0 ) decays to zero if

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

239

 ⊂ Ω− and M = 0. Fig. 4. The dynamics of (4.1) in case D

λ  σ [−
, D], whereas it approximates θ[λ,D,0] (x) if λ > σ [−
, D]. It should be noted that the trivial steady-state u = 0 of (4.1) is linearly stable if λ  σ [−
, D], while it is linearly unstable if λ > σ [−
, D]. Hence, the stability lost by u = 0 as λ crosses σ [−
, D] is gained by the positive solution θ[λ,D,0] bifurcating from u = 0 at σ [−
, D], in complete agreement with the exchange stability principle of Crandall and Rabinowitz [20].

 ⊂ Ω− 4.2. The case M > 0 and D In this case, as an immediate consequence from Theorem 3.5, the next result holds true, because sufficiently large positive constants are supersolutions of (4.3).  ⊂ Ω− , M > 0, and f satisfies (Af ) and (Ag). Then, for each T HEOREM 4.4. Suppose D λ ∈ R, (4.3) possesses a unique positive solution. Moreover, it is unique and strongly positive if it exists, and if we denote it by θ[λ,D,M] and u[λ,D,M] (x, t; u0 ) stands for the unique solution of (4.1), then lim u[λ,D,M] (·, t; u0 ) = θ[λ,D,M]

t↑∞

  . in C D

(4.12)

Further, thanks to Lemma 3.6, the following comparison result holds true.  ⊂ Ω− , M > 0, and f satisfies (Af ) and (Ag). Let u > 0 (resp. L EMMA 4.5. Suppose D u¯ > 0) be a strict subsolution (resp. supersolution) of (4.3). Then u ≪ θ[λ,D,M]

(resp. θ[λ,D,M] ≪ u). ¯

240

J. López-Gómez

Consequently, the estimates 0 < M1  M2 < ∞,

−∞ < λ1  λ2 < ∞,

M 2 − M1 + λ 2 − λ 1 > 0

imply θ[λ1 ,D,M1 ] ≪ θ[λ2 ,D,M2 ] . Moreover, for each λ > σ [−
, D] and M > 0, θ[λ,D,0] ≪ θ[λ,D,M] . As a consequence from Lemma 4.5, the point-wise limit θ[λ,D,∞] := lim θ[λ,D,M] M↑∞

in D

(4.13)

is well defined, though, without any further assumptions on f , it might be everywhere infinity. In the next section we shall see that (4.13) provides us with the minimal positive solution of (4.4) if, in addition, f condition (Ah). Actually, (Ah) is not only sufficient but also necessary for that, though this issue is outside the scope of this work. The next result shows the structural stability of model (4.1) under perturbations of the parameter M ∈ [0, ∞).  ⊂ Ω− and f satisfies (Af ), (Ag). Then P ROPOSITION 4.6. Suppose D
0 if λ  σ [−
, D], lim θ[λ,D,M] = M↓0 θ[λ,D,0] if λ > σ [−
, D]. P ROOF. As a consequence from Lemma 4.5, the point-wise limit Θ0 := lim θ[λ,D,M] M↓0

is well defined. Moreover, by the Schauder theory (e.g., [33]), it is easy to see that Θ0 provides us with a classical solution of (4.3) for M = 0. Necessarily, Θ0
0. Thus, thanks to Theorem 4.2, Θ0 = 0 if λ  σ [−
, D], while, in case λ > σ [−
, D], it follows from Lemma 4.5 that Θ0
θ[λ,D,0] and therefore, by the uniqueness of the positive solution, Θ0 = θ[λ,D,0] , which concludes the proof.  Thanks to Proposition 4.6, the curve of maximal nonnegative solutions of (4.3) for M = 0 perturbs into the curve of positive solutions of (4.3) when M > 0 separates away from M = 0. In Figure 5 we have fixed M > 0 and represented the curve λ → θ[λ,D,M] (x) for a generic x ∈ D, as well as the flow of (4.1). As M ↓ 0 the curve approximates 0 if λ  σ [−
, D], and θ[λ,D,0] (x) if λ > σ [−
, D]. Figure 5 should be compared with Figure 4; the dashed lines represent the nonnegative solutions of (4.3) with M = 0.

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

241

 ⊂ Ω− and M > 0. Fig. 5. The dynamics of (4.1) in case D

 ⊂ Ω− 4.3. The radial case M = ∞ and D = BR (x0 ) with D Subsequently, given x0 ∈ RN and R > 0, BR (x0 ) stands for the open ball  BR (x0 ) := x ∈ RN : |x − x0 | < R .

The main result of this section is the following proposition.  ⊂ Ω− , where D := P ROPOSITION 4.7. Pick λ ∈ R, x0 ∈ Ω− and R > 0 such that D BR (x0 ), and suppose f satisfies (Af ) and (Ag). Consider the auxiliary function H (u) := aL,D g(u)u − λu,

u ∈ [0, ∞),

(4.14)

where aL,D is given by (4.6), and let uλ denote the unique positive zero of H (u). Suppose, in addition, that I (u) :=



u

∞ ' s u

H (z) dz

(−1/2

ds < ∞

for each u > uλ ,

(4.15)

and lim I (u) = 0.

u↑∞

(4.16)

Then, the point-wise limit (4.13) is finite in D, and it provides us with the minimal positive solution of (4.4). P ROOF. Thanks to Theorem 4.4, for each M > 0, (4.3) has a unique positive solution, which has been already denoted by θ[λ,D,M] . By (Ag), θ[λ,D,M] provides us with a positive

242

J. López-Gómez

subsolution of the auxiliary problem
−
u = −H (u) in D = BR (x0 ), u=M

(4.17)

on ∂D.

Thanks again to Theorem 4.4, (4.17) possesses a unique positive solution. Let denote it ΘM . By Lemma 4.5, we have θ[λ,D,M]  ΘM

for each M > 0,

(4.18)

and ΘM1 ≪ ΘM2

if 0 < M1 < M2 .

Thus, the point-wise limit Θ∞ := lim ΘM

(4.19)

M↑∞

is well defined in D. Consequently, due to (4.18), in order to prove that the point-wise limit (4.13) is finite in D it suffices to show that Θ∞ is finite in D. Since (4.17) is invariant by rotations, ΘM must be radially symmetric for each M > 0. Hence, ΘM (x) = ΨM (r),

r := |x − x0 |, x ∈ D,

where ΨM is the unique positive solution of 

 ψ ′′ (r) + N r−1 ψ ′ (r) = H ψ(r) , ψ ′ (0) = 0,

0 < r < R,

ψ(R) = M.

(4.20)

Indeed, the function ψ := 0 is a subsolution of (4.20) and ψ¯ := C is a supersolution of (4.20) for each sufficiently large C > M. Thus, (4.20) has a positive solution, since ¯ necessarily unique, because otherwise we would contradict the uniqueness of the ψ < ψ; positive solution of (4.17). Throughout the remaining of the proof, without lost of generality, we may assume that M > uλ . For this choice, as u := uλ is a positive strict subsolution of (4.17), we find from Lemma 4.5 that uλ < ΘM (x) = ΨM (r) and, consequently,

 H ΨM (r) > 0

 for each x ∈ D

for each r ∈ [0, R],

(4.21)

(4.22)

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

243

since H (z) > 0 for each z > uλ . Also, for each z > uλ , H ′ (z) = aL,D g(z) + aL,D zg ′ (z) − λ

> aL,D g(uλ ) − λ + aL,D zg ′ (z) = aL,D zg ′ (z) > 0

since H (uλ ) = 0. Thus, H is increasing in (uλ , ∞). Now, multiplying the ψ-differential equation by r N −1 and rearranging terms gives

N −1 ′

 ′ r ΨM (r) = r N −1 H ΨM (r) ,

0 < r < R.

(4.23)

Hence, integrating (4.23) in (0, r), yields to ′ ΨM (r) = r 1−N



0

r

 s N −1 H ΨM (s) ds > 0,

r ∈ (0, R),

(4.24)

where we have used (4.22), which, in particular, shows that r → ΨM (r) is increasing, as well as r → H (ΨM (r)). Thus, it follows from (4.24) that

 ′ ΨM (r)  r 1−N H ΨM (r)



0

r

s N −1 ds =

 r H ΨM (r) , N

0 < r < R.

(4.25)

Now, note that (4.25) gives

  N −1 ′ N −1 ′′ ′′ ΨM (r)  ΨM H ΨM (r) H ΨM (r) = ΨM (r) + (r) + r N

and hence,

′′ ΨM (r)


ΨM (r) , N

0 < r < R.

′
0, we find that Similarly, since ΨM

 ′′ ΨM (r)  H ΨM (r)

and therefore,

 ΨM (r) ′′ H ΨM (r)
ΨM , (r)
N

0 < r < R.

(4.26)

′ (r) an integrate in (0, r) to obtain We now multiply (4.26) by ΨM

2



ΨM (r)

ΨM (0)

% ′ &2 2 H (z) dz
ΨM (r)
N



ΨM (r)

ΨM (0)

H (z) dz,

0 < r < R.

(4.27)

244

J. López-Gómez

′ and integrating Thus, taking the square root of the reciprocal of (4.27), multiplying by ΨM the resulting expression in (r, R) shows that

1 √ 2



M ΨM (r)

'

u

H ΨM (0)

(−1/2

du  R − r 

.

N 2



M

ΨM (r)

'

u

ΨM (0)

H

(−1/2

du (4.28)

for each r ∈ [0, R). Now pick r ∈ [0, R). Then, for each M > uλ , we have that ΨM (r)
ΨM (0) and hence, for each u > ΨM (r), 

u

H


ΨM (0)



u

H.

ΨM (r)

Thus, it follows from the second inequality of (4.28) that 0 0, θ[λ,D,M]  Θ∞

in BR−ε (x0 )

and hence, by Shauder’s estimates, there is a constant C(ε) > 0 such that θ[λ,D,M] C 2+µ (BR−2ε (x0 ))  C(ε)

for each M > 0.

Thus, combining the compactness of (−
)−1 with the uniqueness of the point-wise limit (4.13), we find that θ[λ,D,∞] ∈ C 2+µ (BR−2ε (x0 )) must be a solution of (1.9) in BR−2ε (x0 ) and that lim θ[λ,D,M] − θ[λ,D,∞] C (BR−2ε (x0 )) = 0.

M↑∞

As this holds true for any sufficiently small ε > 0, θ[λ,D,∞] must solve (4.4). Actually, it is the minimal positive solution of (4.4). In particular, Θ∞ is the minimal positive solution of (4.30). Indeed, let L be any positive solution of (4.4). Then, for each M > 0, there exists a constant C > M and a sufficiently large n ∈ N such that θ[λ,D,M]  C  L in D \ BR− 1 (x0 ).

(4.31)

n

Thanks to Lemma 4.5, (4.31) implies θ[λ,D,M]  θ[λ,Dn ,C]  L

in Dn := BR− 1 (x0 ) n

and, consequently, for each M > 0, θ[λ,D,M]  L in D. Finally, passing to the limit as M ↑ ∞ in (4.32) we find that θ[λ,D,∞]  L,

(4.32)

246

J. López-Gómez

which shows the minimality of θ[λ,D,∞] and concludes the proof.



 ⊂ Ω− 4.4. The case M = ∞ and D ⊂ Ω− with D As a consequence from Proposition 4.7 the following result is satisfied for a general domain  ⊂ Ω− . D such that D  ⊂ Ω− and f satisfies (Af ), (Ag) and (Ah). Then, for each P ROPOSITION 4.8. Suppose D λ ∈ R, the point-wise limit (4.13) is finite in D and it provides us with the minimal positive solution of the singular problem (4.4). Moreover, for any positive solution u ∈ C 2 (D) ∩  of (1.9) in D one has that C(D) u  θ[λ,D,∞]

in D.

(4.33)

R (x0 ) ⊂ D, and set, for each M > 0, P ROOF. Pick x0 ∈ D and R > 0 such that B αM := max θ[λ,D,M] . ∂BR (x0 )

Then, thanks to Lemma 4.5, for each M > 0 we have that θ[λ,D,M] ≪ θ[λ,BR (x0 ),αM +1]

in BR (x0 ),

and hence, by Proposition 4.7, θ[λ,D,M] ≪ θ[λ,BR (x0 ),∞] < ∞ in BR (x0 ). Therefore, passing to the limit as M ↑ ∞, shows that θ[λ,D,∞]  θ[λ,BR (x0 ),∞] < ∞

in BR (x0 ).

(4.34)

As (4.34) holds true in a small ball around each point x0 ∈ D, θ[λ,D,∞] must be finite in D. Moreover, for each compact subset K ⊂ D there exists a constant C(K) > 0 such that, for each M > 0, θ[λ,D,M]  C(K)

in K.

From these a priori estimates, adapting the last steps of the proof of Proposition 4.7, it is easy to see that θ[λ,D,∞] ∈ C 2+µ (D) must be the minimal positive solution of the singular problem (4.4).  be a positive solution of (1.9) and set To prove (4.33), let u ∈ C 2 (D) ∩ C(D) α := max u. ∂D

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

247

Then, for each M > α, u < M = θ[λ,D,M]

in ∂D,

and hence, thanks to Lemma 4.5, we find that u ≪ θ[λ,D,M]  θ[λ,D,∞]

in D, 

which concludes the proof. 4.5. The general case M ∈ (0, ∞] and D ⊂ Ω−

The next result extends Theorem 4.4 to cover the general case when D ⊂ Ω− . Note that, in this case, a(x) might vanish on some piece of ∂D. T HEOREM 4.9. Suppose D ⊂ Ω− , M > 0, and f satisfies (Af ), (Ag) and (Ah). Then, (4.3) possesses a unique positive solution for each λ ∈ R. Moreover, it is strongly positive and, if we denote it by θ[λ,D,M] , then the unique solution of (4.1), u[λ,D,M] (x, t; u0 ), satisfies (4.12). P ROOF. Subsequently, for each sufficiently large n ∈ N, say n
n0 , we consider the subdomain of D defined by


1 . (4.35) Dn := x ∈ D: dist(x, ∂D) > n Then, for each n
n0 , n ⊂ Dn+1 ⊂ D ⊂ Ω− , D

aL,Dn := min a > 0 n D

(4.36)

and D=

∞ *

Dn .

(4.37)

n=n0

Due to (4.36), Theorem 4.4 guarantees that θ[λ,Dn ,M] is well defined; recall that θ[λ,Dn ,M] stands for the unique positive solution of
−
u = λu − af (·, u)u in Dn , u=M on ∂Dn . Also, thanks to Proposition 4.8, for each n
n0 , the point-wise limit θ[λ,Dn ,∞] := lim θ[λ,Dn ,M] M↑∞

248

J. López-Gómez

Fig. 6. The positive solutions of (4.38) for large M.

is finite and it provides us with the minimal positive solution of the singular problem
−
u = λu − af (·, u)u in Dn , u=∞ on ∂Dn . Moreover, for each n
n0 and m
1 (see Figure 6), θ[λ,Dn+m ,M] ≪ θ[λ,Dn+m ,∞] ≪ θ[λ,Dn ,∞]

in Dn .

(4.38)

In particular, due to (4.37), for each compact subset K ⊂ D there exists a constant C(K) > 0 and an integer nK ∈ N such that, for each M > 0 and n
nK , θ[λ,Dn ,M]  C(K)

in K.

(4.39)

From (4.39), a diagonal argument combined with the compactness of (−
)−1 , shows the existence of a positive solution of (4.3), denoted by θ[λ,D,M] . As the positive solution θ[λ,D,M] itself provides us with a positive supersolution of (4.3), the remaining assertions of the theorem follow as direct consequences from Theorem 3.5.  Also, thanks to Lemma 3.6, the following counterpart of Lemma 4.5 holds true. L EMMA 4.10. Suppose D ⊂ Ω− , M > 0, and f satisfies (Af ), (Ag) and (Ah). Let u > 0 (resp. u¯ > 0) be a strict subsolution (resp. supersolution) of (4.3). Then u ≪ θ[λ,D,M]

(resp. θ[λ,D,M] ≪ u). ¯

Consequently, the estimates 0 < M1  M2 < ∞,

−∞ < λ1  λ2 < ∞,

M 2 − M1 + λ 2 − λ 1 > 0

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

249

imply θ[λ1 ,D,M1 ] ≪ θ[λ2 ,D,M2 ] . Now, we can give the main existence result of positive solutions for the singular problem (4.4) in case D ⊂ Ω− . It should be noted that Theorem 2.1 is a direct consequence from it by making the choice D = Ω− . T HEOREM 4.11. Suppose D ⊂ Ω− and f satisfies (Af ), (Ag) and (Ah). Then (4.4) posmax sesses a minimal an a maximal positive solution, denoted by Lmin [λ,D] and L[λ,D] , respectively, i.e., any other positive solution L of (4.4) satisfies max Lmin [λ,D]  L  L[λ,D] .

Moreover, Lmin [λ,D] = θ[λ,D,∞] := lim θ[λ,D,M]

(4.40)

min Lmax [λ,D] = lim θ[λ,Dn ,∞] = lim L[λ,Dn ] ,

(4.41)

M↑∞

and n↑∞

n↑∞

where Dn , n
1, is the subdomain of D defined by (4.35). P ROOF. Thanks to Lemma 4.10, the point-wise limit (4.40) is well defined. Now, we shall see that it is finite. Pick x0 ∈ D and R > 0 such that R (x0 ) ⊂ D ⊂ Ω− B and set αM := max θ[λ,D,M] , ∂BR (x0 )

M > 0.

Thanks to Lemma 4.10, θ[λ,D,M] ≪ θ[λ,BR (x0 ),αM +1]

in BR (x0 )

and therefore, for each M > 0, θ[λ,D,M] ≪ θ[λ,BR (x0 ),∞]

in BR (x0 ).

As this argument is valid for any x0 ∈ D and, thanks to Proposition 4.7, the point-wise limit θ[λ,BR (x0 ),∞] is finite in BR (x0 ), for each compact subset K ⊂ D there exists a constant C(K) > 0 such that, for each M > 0, θ[λ,D,M]  C(K)

in K.

250

J. López-Gómez

This shows that the point-wise limit (4.40) is finite in D. Further, combining the monotonicity of the involved sequences with the compactness of (−
)−1 , it is easy to see that θ[λ,D,∞] ∈ C 2+µ (D) provides us with a positive solution of (4.4). Actually, it is the minimal positive solution. Indeed, let L be any positive solution of (4.4). Then, for each M > 0, there exists a constant C > 0 and a sufficiently large n ∈ N such that θ[λ,D,M]  C  L

in D \ Dn .

(4.42)

Also, thanks to Lemma 4.5, (4.42) implies θ[λ,D,M]  θ[λ,Dn ,C]  L in Dn and therefore, θ[λ,D,M]  L

in D.

(4.43)

Consequently, passing to the limit as M ↑ ∞ in (4.43) we find that θ[λ,D,∞]  L which shows the minimality of θ[λ,D,∞] as a positive solution of the singular problem (4.4). Similarly, thanks to (4.38) and (4.39), the point-wise limit (4.41) is well defined in D and it provides us with a positive solution of (4.4). To show its maximality, suppose L is any positive solution of (4.4). Then, for each n
n0 , there exists M > 0 such that L|Dn provides us with a positive subsolution of (4.3), and therefore, thanks to Lemma 4.5, L  θ[λ,Dn ,M]  θ[λ,Dn ,∞]

in Dn .

Consequently, passing to the limit as n ↑ ∞ gives L  Lmax [λ,D] , which concludes the proof.



max In Figure 7 we have sketched the construction of Lmin [λ,D] and L[λ,D] carried out in the proof of Theorem 4.11. Although we conjecture that, under the general assumptions of this section, max Lmin [λ,D] = L[λ,D] ,

i.e., the positive solution of (4.4) is unique, we could not get a proof of this fact, except – essentially – in the case described by Theorem 2.7.

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

251

min Fig. 7. The construction of Lmax [λ,D] and L[λ,D] .

5. Proofs of Theorems 2.1–2.3 Theorem 2.1 is a direct consequence from Theorem 4.11. Actually, thanks to Theorem 4.11, for each λ ∈ R the following relation is satisfied Lmin [λ,Ω− ] = lim θ[λ,Ω− ,M] ,

(5.1)

M↑∞

where θ[λ,Ω− ,M] stands for the unique positive solution of (2.1) with D = Ω− . Moreover, min Lmax [λ,Ω− ] = lim L[λ,Ω n ] , n↑∞

(5.2)

−

where, for any sufficiently large n ∈ N, n Ω−


1 . := x ∈ Ω− : dist(x, ∂Ω− ) > n

Now, we shall begin the proof of Theorem 2.3. First, we will characterize the existence of positive solutions of


−
u = λu − af (·, u)u u=0

in Ω, on ∂Ω.

(5.3)

Suppose u is a positive solution of (5.3). Then, thanks to Lemma 3.4, u ≫ 0 and % & λ = σ −
+ af (·, u), Ω .

(5.4)

252

J. López-Gómez

In particular, af (·, u) > 0 in Ω, and hence, by the monotonicity of the principal eigenvalue with respect to the potential, (5.4) implies % & λ = σ −
+ af (·, u), Ω > σ [−
, Ω] = σ0 .

Also, by the monotonicity of the principal eigenvalue with respect to the domain, it is apparent that % & % & λ = σ −
+ af (·, u), Ω < σ −
+ af (·, u), Ω0,1 = σ [−
, Ω0,1 ] = σ1

since a = 0 in Ω0,1 . Therefore, condition σ0 < λ < σ1

(5.5)

is necessary for the existence of a positive solution of (5.3). Note that σ0 = σ [−
, Ω] < σ [−
, Ω0,1 ] = σ1 by the monotonicity of the principal eigenvalue with respect to the domain. Now, we will show that (5.5) is not only necessary but also sufficient for the existence of a positive solution of (5.3). Suppose (5.5). Then, due to Theorem 3.5, to show the existence of a positive solution it suffices to construct a positive supersolution of (5.3). Actually, thanks to Theorem 3.5, the existence of the positive supersolution itself entails its own uniqueness. It should be noted that, thanks to Theorem 3.5, (5.3) cannot admit a positive supersolution if λ
σ1 . To construct the supersolution we proceed as follows. For each j ∈ {1, 2} and sufficiently small δ > 0 consider the open δ-neighborhoods  Ωδ,j := x ∈ Ω: dist(x, Ω0,j ) < δ ,

which have been represented in Figure 8. Note that Ωδ,1 consists of Ω0,1 , Γ1 and the set 0,2 and the set of of points x ∈ Ω− such that dist(x, Γ1 ) < δ. Similarly, Ωδ,2 consists of Ω points x ∈ Ω− such that dist(x, Γ2 ) < δ. By the continuous dependence of the principal eigenvalues with respect to the domain (e.g., [43], Theorem 4.2), we have that lim σ [−
, Ωδ,j ] = σ [−
, Ω0,j ] = σj , δ↓0

j ∈ {1, 2}.

Thus, by the monotonicity of the principal eigenvalues with respect to the domains, it is apparent that, for each sufficiently small δ > 0, σ0 < λ < σ [−
, Ωδ,1 ] < σ1 < σ [−
, Ωδ,2 ] < σ2 .

(5.6)

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

253

Fig. 8. The δ-neighborhoods Ωδ,1 and Ωδ,2 .

Fix λ ∈ (σ0 , σ1 ) and pick one of these values of δ. Now, for each j ∈ {1, 2}, let ϕδ,j ≫ 0 denote a principal eigenfunction associated with σ [−
, Ωδ,j ] – unique up to a multiplicative constant – and consider the function Φ defined through    ϕδ,1 Φ := ϕδ,2   ϕ−

δ/2,1 , in Ω δ/2,2 , in Ω 

δ/2,1 ∪ Ω δ/2,2 , in Ω \ Ω

(5.7)

where ϕ− is any smooth extension of ϕδ,1 ∨ ϕδ,2 to




 δ δ/2,2 = x ∈ Ω− : dist(x, ∂Ω− ) > δ/2,1 ∪ Ω Ω\ Ω 2 that it is positive and bounded away from zero. Note that ϕ− exists, since ϕδ,j is positive and bounded away from zero on Ω− ∩ ∂Ωδ/2,j , j ∈ {1, 2}. Figure 9 shows a genuine profile of Φ. We claim that the function u¯ := κΦ is a supersolution of (5.3) for each sufficiently large κ > 1. Indeed, κΦ = 0 on ∂Ω

(5.8)

254

J. López-Gómez

Fig. 9. The profile of the supersolution element Φ.

by construction. Moreover, for each j ∈ {1, 2}, we have that −
(κΦ)
λκΦ − af (·, κΦ)κΦ

in Ωδ/2,j

if and only if κσ [−
, Ωδ,j ]ϕδ,j
λκϕδ,j − af (·, κϕδ,j )κϕδ,j

in Ωδ/2,j

or equivalently, af (·, κϕδ,j )
λ − σ [−
, Ωδ,j ]

in Ωδ/2,j

which holds true for every κ > 0 since, due to (5.6), af (·, κϕδ,j )
0 > λ − σ [−
, Ωδ,j ]

in Ωδ,j .

Further, we have that −
(κΦ)
λκΦ − af (·, κΦ)κΦ if and only if −
ϕ−
λ − af (·, κϕ− ) ϕ−



δ/2,2 δ/2,1 ∪ Ω in Ω \ Ω

 δ/2,1 ∪ Ω δ/2,2 in Ω \ Ω

which holds true for all sufficiently large κ > 1, by (Ag), since a and ϕ− are positive and bounded away from zero in 

δ/2,2 ⊂⊂ Ω− . δ/2,1 ∪ Ω Ω\ Ω

Therefore, (5.8) provides us with a supersolution of (5.3) for sufficiently large κ > 1. Consequently, by Theorem 3.5, (5.3) possesses a positive solution if and only if condition (5.5) is satisfied. Moreover, it is strongly positive and unique. Subsequently, the unique positive solution of (5.3) will be denoted by θ (λ) = θ[λ,Ω] = θ[λ,Ω,0] .

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

255

Now, the proof of Proposition 4.3 can be easily adapted to prove (2.6) and to show that the map (2.8) in point-wise increasing. Actually, θ[λ,Ω] ≪ θ[µ,Ω]

if σ0 < λ < µ < σ1 ,

which can be derived from Lemma 3.6. So, the technical details of the proofs of these features, by repetitive, will be omitted here in. Now, we will prove the following property lim θ[λ,Ω] = ∞ uniformly in compact subsets of Ω0,1 .

(5.9)

λ↑σ1

Pick λ1 ∈ (σ0 , σ1 ) and consider η > 0 such that θ[λ1 ,Ω] > ηϕ0,1

in Ω0,1 ,

(5.10)

where ϕ0,1 ≫ 0 is a principal eigenfunction associated with σ1 = σ [−
, Ω0,1 ]. By differentiating the realization of (5.3) at θ (λ) with respect to λ, particularizing the result at θ (λ) = θ[λ,Ω] , and rearranging terms gives 



  dθ −
+ a ∂f ∂u ·, θ(λ) θ (λ) + af ·, θ(λ) − λ dλ (λ) = θ (λ) dθ dλ (λ) = 0

in Ω,

(5.11)

on ∂Ω.

It should be noted that (5.11) follows straight ahead from the proof of the differentiability of the mapping λ → θ (λ) := θ[λ,Ω] through the adaptation of the proof of Proposition 4.3. Since a and

 ∂f ·, θ(λ) θ (λ) > 0 in Ω ∂u

%

 & λ = σ −
+ af ·, θ(λ) , Ω ,

we find, from the monotonicity of the principal eigenvalue with respect to the potential, that ' ( 

 ∂f ·, θ(λ) θ (λ) + af ·, θ(λ) − λ, Ω σ −
+ a ∂u %

 & > σ −
+ af ·, θ(λ) − λ, Ω = 0

and therefore, thanks to Theorem 3.1, the differential operator −
+ a



 ∂f ·, θ(λ) θ (λ) + af ·, θ(λ) − λ ∂u

256

J. López-Gómez

satisfies the strong maximum principle in Ω under Dirichlet boundary conditions. In pardθ ticular, it follows from (5.11) that dλ (λ) ≫ 0 and, consequently, λ → θ (λ) is strongly increasing. Now, due to (5.10), for each λ ∈ [λ1 , σ1 ), we have that θ[λ,Ω] > ηϕ0,1

in Ω0,1 ,

and hence, (5.11) gives  dθ (λ) > ηϕ0,1 (−
− λ) dλ dθ dλ (λ) > 0

since

dθ dλ (λ)|Γ1

in Ω0,1 ,

(5.12)

on ∂Ω0,1

> 0. Moreover, for each λ ∈ [λ1 , σ1 ),

σ [−
− λ, Ω0,1 ] = σ1 − λ > 0 and therefore, thanks again to Theorem 3.1, we find from (5.12) that dθ (λ) ≫ Ψλ dλ

in Ω0,1 ,

(5.13)

where Ψλ is the unique solution of
(−
− λ)Ψλ = ηϕ0,1 in Ω0,1 , Ψλ = 0 on ∂Ω0,1 . A direct calculation shows that Ψλ =

η ϕ0,1 , σ1 − λ

and hence, (5.13) gives lim

λ↑σ1

dθ (λ) = ∞ dλ

uniformly in compact subsets of Ω0,1 .

Consequently, (5.9) holds true. Before ending the proof of Theorem 2.3, we shall begin the proof of Theorem 2.2. Subsequently, for each M > 0, we consider the following nonlinear boundary value problem
−
u = λu − af (·, u)u in D, (5.14) u=M on ∂D, where 0,1 = Ω− ∪ Ω 0,2 . D := Ω \ Ω

(5.15)

The notation (5.15) will be maintain in the remaining of this section. The following result is satisfied.

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

257

L EMMA 5.1. Problem (5.14) possesses a positive solution if and only if λ < σ2 . Moreover, it is unique if it exists and if we denote it by θ[λ,D,M] , then 0 ≪ θ[λ1 ,D,M1 ] ≪ θ[λ2 ,D,M2 ]

in D

if −∞ < λ1  λ2 < σ2 ,

0 < M1  M2 ,

λ2 − λ1 + M2 − M1 > 0.

¯ Also, for any λ < σ2 and any positive strong subsolution (resp. supersolution) u (resp. u) of (5.14), one has that u ≪ θ[λ,D,M] (resp. u¯ ≫ θ[λ,D,M] ). Furthermore, for each M > 0, λ < σ2 and u0 > 0,

  , lim u[λ,D,M] (·, t; u0 ) = θ[λ,D,M] in C D t↑∞

where u[λ,D,M] (x, t; u0 ) stands for the unique solution of the parabolic problem  ∂u   ∂t −
u = λu − af (·, u)u in D × (0, ∞), u=M on ∂D × (0, ∞),   in D. u(·, 0) = u0 P ROOF. For any sufficiently small δ > 0, consider the δ-neighborhood of Γ1  Γ1,δ := Γ1 + Bδ (0) = x ∈ Ω: dist(x, Γ1 ) < δ

(5.16)

(5.17)

and the open subdomain of Ω defined by Ω1,δ := D ∪ Γ1,δ ;

(5.18)

δ > 0 must be chosen sufficiently small so that ∂Ω1,δ ⊂ Ω0,1

and σ [−
, Γ1,δ ] > σ2 = σ [−
, Ω0,2 ].

(5.19)

The second relation of (5.19) can be got since the Lebesgue measure of Γ1,δ decays to zero as δ ↓ 0, by Faber–Krahn inequality (cf. [43], if necessary). Suppose u > 0 is a solution of (5.14). Then, thanks to Lemma 3.4, u ≫ 0. Moreover, since u is a positive strict supersolution of −
− λ + af (·, u) in D under homogeneous Dirichlet boundary conditions, it follows from Theorem 3.1 that % & % & 0 < σ −
− λ + af (·, u), D  σ −
− λ + af (·, u), Ω0,2 = σ2 − λ

by the monotonicity of the principal eigenvalue with respect to the domain. Note that a = 0 in Ω0,2 . Thus, λ < σ2 is necessary for the existence of a positive solution of (5.14). Consequently, in the remaining of the proof we suppose λ < σ2 .

258

J. López-Gómez

Thanks to Theorem 3.5 and Lemma 3.6, in order to complete the proof of Lemma 5.1 it suffices to construct a positive supersolution for (5.14). Let ϕ1,δ ≫ 0 be a principal eigenfunction associated with σ [−
, Γ1,δ ]. Note that, due to (5.19), λ < σ2 < σ [−
, Γ1,δ ].

(5.20)

Now, reduce δ > 0, if necessary, so that the δ-neighborhood of Ω0,2 defined by

satisfy

 Ωδ,2 := Ω0,2 + Bδ (0) = x ∈ Ω: dist(x, Ω0,2 ) < δ λ < σ [−
, Ωδ,2 ] < σ2 ,

(5.21)

and pick a principal eigenfunction associated with σ [−
, Ωδ,2 ], say ϕδ,2 ≫ 0. Fix a sufficiently small δ > 0 satisfying the previous requirements and consider the function Ψ defined by  ϕ1,δ    Ψ := ϕδ,2    ϕ−

 1,δ \ x ∈ D: dist(x, Γ1 ) > δ , in Ω 2 δ , in Ω 2 ,2  in x ∈ Ω− : dist(x, ∂Ω− ) > 2δ ,

(5.22)

where ϕ− is any positive smooth extension, bounded away from zero, of ϕ1,δ ∨ ϕδ,2 to


δ . Ω−,δ := x ∈ Ω− : dist(x, ∂Ω− ) > 2

We claim that the function u¯ := κΨ is a positive supersolution of (5.14) in D if κ > 1 is sufficiently large. Indeed, by construction, u¯ ≫ 0, and κΨ > M on Γ1 = ∂D for sufficiently large κ > 1 since min ϕ1,δ > 0. Γ1

Moreover, in the set


δ x ∈ Ω− : dist(x, Γ1 )  2 we have that −
(κΨ )
λκΨ − af (·, κΨ )κΨ

(5.23)

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

259

if and only if κσ [−
, Γ1,δ ]ϕ1,δ
λκϕ1,δ − af (·, κϕ1,δ )κϕ1,δ , or equivalently, af (·, κϕ1,δ )
λ − σ [−
, Γ1,δ ] which is true because, due to (5.20), af (·, κϕ1,δ )
0 > λ − σ [−
, Γ1,δ ]. Similarly, in Ωδ/2,2 , (5.23) is satisfied if and only if af (·, κϕδ,2 )
λ − σ [−
, Ωδ,2 ], which holds true by (5.21). Finally, in Ω−,δ , (5.23) is satisfied if and only if af (·, κϕ− )
λ +


ϕ− , ϕ−

which holds true for sufficiently large κ > 1, since ϕ− and a are positive and bounded away from zero. Consequently, u¯ = κΨ provides us with a positive supersolution of (5.14) for any sufficiently large κ > 1, which ends the proof.  Thanks to Lemma 5.1, the point-wise limit θ[λ,D,∞] := lim θ[λ,D,M]

(5.24)

M↑∞

is well defined in D for each λ < σ2 . We claim that it is finite everywhere in D. Indeed, setting 0 / αM := max M, max θ[λ,D,M] + 1, M > 0, Γ2

we have that θ[λ,D,M] |∂Ω− < αM ,

M > 0,

and hence, θ[λ,D,M] |Ω− provides us with a positive strong subsolution of


−
u = λu − af (·, u)u u = αM

in Ω− , on ∂Ω− .

Thus, thanks to Lemma 4.10 and Theorem 4.11, we have that, for each M > 0, θ[λ,D,M] ≪ θ[λ,Ω− ,αM ] ≪ Lmin [λ,Ω− ]

in Ω− ,

(5.25)

260

J. López-Gómez

where Lmin [λ,Ω− ] stands for the minimal large solution of (1.9) in Ω− . Therefore, passing to the limit as M ↑ ∞ gives θ[λ,D,∞]  Lmin [λ,Ω− ]

in Ω− .

(5.26)

0,2 we can use the This shows that (5.24) is finite in Ω− . To prove that it is finite in Ω following result. L EMMA 5.2. For each sufficiently large n ∈ N, say n
n0 , let Dn denote the open set defined in (4.35), where D is given by (5.15), and, for each M > 0, consider the boundary value problem


−
u = λu − af (·, u)u u=M

in Dn , on ∂Dn .

(5.27)

Then, (5.27) possesses a positive solution if and only if λ < σ2 . Moreover, it is unique if it exists and if we denote it by θ[λ,Dn ,M] , then, for each n
n0 , 0 ≪ θ[λ1 ,Dn ,M1 ] ≪ θ[λ2 ,Dn ,M2 ]

in D

provided −∞ < λ1  λ2 < σ2 ,

0 < M1  M2 ,

λ2 − λ1 + M2 − M1 > 0.

Moreover, for any λ < σ2 and any positive strict subsolution (resp. supersolution) u (resp. u) ¯ of (5.27) one has that u ≪ θ[λ,Dn ,M] (resp. u¯ ≫ θ[λ,Dn ,M] ). Furthermore, for each M > 0, λ < σ2 , and u0 > 0, lim u[λ,Dn ,M] (·, t; u0 ) = θ[λ,Dn ,M] ,

t↑∞

where u[λ,Dn ,M] (x, t; u0 ) stands for the unique solution of the parabolic problem  ∂u   ∂t −
u = λu − af (·, u)u u=M   u(·, 0) = u0

in Dn × (0, ∞),

on ∂Dn × (0, ∞), in Dn .

The proof of Lemma 5.2 is easier than the proof of Lemma 5.1 since a is positive and bounded away from zero on ∂Dn , which simplifies the construction of the basic supersolution pattern Ψ . Therefore, its technical details are omitted here in by repetitive. Now, fix n
n0 . As ∂Dn is a compact subset of Ω− , due to (5.26), there exists a constant C(n) > 0 such that, for each M > 0, θ[λ,D,M]  C(n)

on ∂Dn .

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

261

Thus, thanks to Lemma 5.2, for each M > 0 we find that θ[λ,D,M]  θ[λ,Dn ,C(n)]

in Dn

and, consequently, θ[λ,D,∞] is finite in Dn . As this argument is valid for each n
n0 and D=

∞ *

Dn ,

(5.28)

n=n0

necessarily θ[λ,D,∞] is finite in D. The previous argument can be easily adapted to show that, for any n
n0 , each of the point-wise limits θ[λ,Dn ,∞] := lim θ[λ,Dn ,M]

(5.29)

M↑∞

is well defined and finite in Dn . Now, adapting the argument of the final part of the proof of Proposition 4.7, it is easy to see that θ[λ,D,∞] ∈ C 2+µ (D) provides us with a positive solution of


−
u = λu − af (·, u)u

u=∞

in D, on ∂D.

(5.30)

Actually, it is the minimal positive solution. Indeed, let L be any positive solution of (5.30). Then, for each M > 0, there exist a constant C > 0 and a sufficiently large n ∈ N such that n . θ[λ,D,M]  C  L in D \ D

(5.31)

Moreover, thanks to Lemma 5.2, (5.31) implies θ[λ,D,M]  θ[λ,Dn ,C]  L

in Dn ,

and consequently, θ[λ,D,M]  L in D.

(5.32)

Therefore, passing to the limit as M ↑ ∞ in (5.32) we find that θ[λ,D,∞]  L, which shows its minimality. Consequently, the solution θ[λ,D,∞] provides us with the large solution Lmin [λ,D] referred to in the statement of Theorem 2.2. Similarly, for each n
n0 , the point-wise limit (5.29) provides us with the minimal positive solution of
−
u = λu − af (·, u)u in Dn , (5.33) u=∞ on ∂Dn ,

262

J. López-Gómez

which will be subsequently denoted by Lmin [λ,Dn ] . We now show that the point-wise limit min Lmax [λ,D] := lim L[λ,Dn ]

(5.34)

n↑∞

is well defined and that, actually, it provides us with the maximal solution of (5.30). Indeed, for each M > 0, n
n0 and m
1, we have that θ[λ,Dn+m ,M]  θ[λ,Dn+m ,∞]  θ[λ,Dn ,∞]

in Dn .

(5.35)

Thus, due to (5.28), for each compact subset K ⊂ D, there exist C(K) > 0 and nK ∈ N such that, for each M > 0 and n
nK , θ[λ,Dn ,M]  C(K)

in K,

and hence, Lmin [λ,Dn ]  C(K)

in K.

(5.36)

From these a priori estimates, a diagonal argument combined with the compactness of (−
)−1 , shows that (5.34) is finite in D and that it provides us with a solution of (5.30). To show the maximality of Lmax [λ,D] , let L be any positive solution of (5.30). Then, for each n
n0 , there exists M > 0 such that L|Dn provides us with a positive subsolution of (5.27), and therefore, thanks to Lemma 5.2, L  θ[λ,Dn ,M]  θ[λ,Dn ,∞]

in Dn .

Consequently, passing to the limit as n ↑ ∞ gives L  Lmax [λ,D] . To conclude the proof of the existence of Theorem 2.2 it remains to shows that λ < σ2 is necessary for the existence of a positive solution of (5.30). Suppose L is a solution of (5.30) for some λ ∈ R. Then, for each n
n0 , L|Dn provides us with a positive strict supersolution of −
+ af (·, L) − λ in Dn under homogeneous Dirichlet boundary conditions, and hence, thanks to Theorem 3.1, & % σ −
+ af (·, L) − λ, Dn > 0.

Therefore, by the monotonicity of the principal eigenvalue with respect to the domain & % & % λ < σ −
+ af (·, L), Dn < σ −
+ af (·, L), Ω0,2 = σ2

since a = 0 in Ω0,2 . Note that, in order to complete the proof of Theorem 2.2, it remains to show (2.5).

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

263

Now, we shall show that lim Lmin [λ,D] = ∞

λ↑σ2

uniformly in compact subsets of Ω0,2 .

(5.37)

To prove (5.37) we need the following result. L EMMA 5.3. Consider the problem
−
u = λu − af (·, u)u in D, u=0 on ∂D,

(5.38)

0,1 . Then, (5.38) possesses a positive solution if and only if where D = Ω \ Ω σ [−
, D] < λ < σ2 .

(5.39)

Moreover, it is unique if it exists and if we denote it by θ[λ,D] , then the mapping

 θ

  , σ [−
, D], σ2 −→ C Ω λ → θ (λ) := θ[λ,D]

(5.40)

is point-wise increasing and of class C 1 . Furthermore, lim θ[λ,D] = ∞ uniformly in compact subsets of Ω0,2 .

λ↑σ2

(5.41)

P ROOF. Suppose u > 0 is a solution of (5.38). Then, by Lemma 3.4, u ≫ 0 and

Thus,

% & λ = σ −
+ af (·, u), D .

% & σ [−
, D] < λ = σ −
+ af (·, u), D < σ [−
, Ω0,2 ] = σ2

since a = 0 in Ω0,2 . Therefore, (5.39) is necessary for the existence of a positive solution. Suppose (5.39). Then, thanks to Lemma 5.1, the problem


−
u = λu − af (·, u)u u=1

in D, on ∂D

possesses a unique positive solution, θ[λ,D,1] . Clearly, u¯ := θ[λ,D,1] is a positive supersolution of (5.38) and consequently, the lemma is a direct consequence from Theorem 3.5 and Lemma 3.6, except for the regularity of the map defined in (5.40) and (5.41), which can be easily obtained by adapting the argument used to prove the regularity of the map (2.8) and the proof of (5.9), respectively. By repetitive, we shall omit the technical details here in. 

264

J. López-Gómez

Now, we are ready to prove (5.37). Thanks to Lemma 5.1, for each M > 0 and λ, µ ∈ R, λ < µ < σ2 , we have that θ[λ,D,M] ≪ θ[µ,D,M] and hence, passing to the limit as M ↑ ∞ gives min Lmin [λ,D]  L[µ,D] .

Thus, the point-wise limit Lσ2 := lim Lmin [λ,D]

(5.42)

λ↑σ2

 Moreover, for each M > 0, we have that is well defined in D. Lmin [λ,D] ≫ θ[λ,D,M] ≫ θ[λ,D] , and therefore, (5.37) follows from (5.41). Subsequently, we shall complete the proof of (2.7), which concludes the proof of Theorem 2.3. For a fixed n
n0 we will consider the open set Dn defined in (4.35) with 0,1 . For each M > 0, let θ[σ1 ,Dn ,M] be the unique positive solution of (5.27) at D=Ω \Ω λ = σ1 . Such solution exists by Lemma 5.2. For each λ ∈ (σ0 , σ1 ), set Mλ := max θ[λ,Ω] . ∂Dn

Then θ[λ,Ω] |∂Dn  Mλ and, thanks to Lemma 5.2, we find that θ[λ,Ω]  θ[σ1 ,Dn ,Mλ ]

in Dn .

Thus, θ[λ,Ω]  Lmin [σ1 ,Dn ]

in Dn

(5.43)

since Lmin [σ1 ,Dn ] = lim θ[σ1 ,Dn ,M] . M↑∞

As the mapping λ → θ[λ,Ω] is increasing and (5.43) holds for each λ ∈ (σ0 , σ1 ) and n
n0 , it is apparent that the point-wise limit Lσ1 := lim θ[λ,Ω] λ↑σ1

in D

(5.44)

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

265

is well defined and finite. Actually, Lσ1 ∈ C 2+µ (D) provides us with a solution of (1.9) in D. In order to prove the identity Lσ1 = Lmin [σ1 ,D] , it suffices to show that lim min θ[λ,Ω] = ∞

(5.45)

λ↑σ1 ∂D

(note that ∂D = Γ1 ). Indeed, suppose (5.45) has been proven. Then, setting mλ := min θ[λ,Ω] , ∂D

λ ∈ (σ0 , σ1 ),

it follows from Lemma 5.1 that θ[λ,Ω]
θ[λ,D,mλ ]

in D.

In particular, for each ε > 0 and λ ∈ (σ1 − ε, σ1 ), θ[λ,Ω]
θ[σ1 −ε,D,mλ ]

in D.

Thus, passing to the limit as λ ↑ σ1 , (5.44) implies Lσ1
Lmin [σ1 −ε,D]

in D

for each ε > 0. This shows that Lσ1 provides us with a large solution of (1.9) in D. Similarly, we have that θ[λ,Ω]  θ[λ,D,max∂D θ[λ,Ω] ]  Lmin [σ1 ,D]

in D

and therefore, Lσ1  Lmin [σ1 ,D]

in D.

Consequently, Lσ1 = Lmin [σ1 ,D]

in D,

and the proof of Theorem 2.3 will be completed if we show (5.45). Note that (5.9) and (5.45) imply lim θ[λ,Ω] = ∞

λ↑σ1

0,1 \ ∂Ω. in Ω

Thus, 1/θ[λ,Ω] , λ ∈ (σ0 , σ1 ), provides us with a decreasing family of continuous functions 0,1 \ ∂Ω, and therefore, it approximates zero uniformly point-wise converging to zero in Ω

266

J. López-Gómez

0,1 \ ∂Ω. Consequently, θ[λ,Ω] approximates ∞ as λ ↑ σ1 uniin compact subsets of Ω 0,1 \ ∂Ω. formly in compact subsets of Ω To prove (5.45) we argue by contradiction. Suppose (5.45) is not satisfied. Then b := min Lσ1 ∈ (0, ∞).

(5.46)

∂D

As ∂D = Γ1 = ∂Ω0,1 \ ∂Ω is of class C 2 , there exist R > 0 and a map Y : ∂D → Ω0,1 such that, for each x ∈ ∂D = Γ1 ,

 BR Y (x) ⊂ Ω0,1 ,

 R Y (x) ∩ ∂Ω = ∅, B

 R Y (x) ∩ ∂D = {x}. B

Moreover, thanks to (5.46), for each λ ∈ (σ0 , σ1 ), there exists xλ ∈ ∂D such that θ[λ,Ω] (xλ ) = min θ[λ,Ω]  min Lσ1 = b. ∂D

∂D

By construction,

 dist xλ , Y (xλ ) = R

for each λ ∈ (σ0 , σ1 ).

Moreover, the manifold Γ1R defined by .  Γ1R = y ∈ Ω0,1 : dist(y, ∂D) = 2R

is a compact subset of Ω0,1 , and hence, thanks to (5.9), lim θ[λ,Ω] = ∞

λ↑σ1

uniformly in Γ1R .

(5.47)

Setting  R Ω0,1 := y ∈ Ω0,1 : dist(y, ∂D) < 2R ,

R , while the other one consists of ∂D. it turns out that Γ1R is one of the components of ∂Ω0,1 R be such that Let x˜λ ∈ Ω 0,1

θ[λ,Ω] (x˜λ ) = min θ[λ,Ω] R Ω 0,1

R . Then and suppose x˜λ ∈ Ω0,1

∇θ[λ,Ω] (x˜λ ) = 0,


θ[λ,Ω] (x˜λ )
0,

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

267

and, since R (−
− λ)θ[λ,Ω] = 0 in Ω0,1 ,

we find that −
θ[λ,Ω] (x˜λ ) = λθ[λ,Ω] (x˜λ ) > 0 for each λ ∈ (σ0 , σ1 ) because λ > 0, which is contradictory. Consequently, for each λ ∈ (σ0 , σ1 ), R x˜λ ∈ ∂Ω0,1 = Γ1 ∪ Γ1R .

Therefore, thanks to (5.47), there exists σˆ 1 ∈ (σ0 , σ1 ) such that min θ[λ,Ω] = min θ[λ,Ω] = θ[λ,Ω] (xλ )  b R Ω 0,1

∂D

for each λ ∈ (σˆ 1 , σ1 ). In particular, θ[λ,Ω] (x)
θ[λ,Ω] (xλ )

 R Y (xλ ) . for each x ∈ B

(5.48)

Now, for each α > 0 and λ ∈ (σˆ 1 , σ1 ), we consider the auxiliary function ψλ defined by 2

ψλ (x) := e−α|x−Y (xλ )| − e−αR

2

 R Y (xλ ) . for each x ∈ B

A direct calculation shows that, for each x ∈ BR (Y (xλ )),

2  

2 2 (−
− λ)ψλ (x) = 2αN − 4α 2 x − Y (xλ ) − λ e−α|x−Y (xλ )| + λe−αR

and hence, there exist α > 0 and ω > 0 such that, for each λ ∈ (σˆ 1 , σ1 ), (−
− λ)ψλ  −ω



 R/2 Y (xλ ) . in AR := BR Y (xλ ) \ B

(5.49)

R/2 (Y (xλ )) Subsequently, we will assume that α has been chosen to satisfy (5.49). Since B is a compact subset of Ω0,1 , it follows from (5.9) that lim

min

λ↑σ1 B R/2 (Y (xλ ))

θ[λ,Ω] = ∞

and hence, setting cλ :=

minBR/2 (Y (xλ )) θ[λ,Ω] − θ[λ,Ω] (xλ ) e−αR

2 /4

− e−αR

2

268

J. López-Gómez

we have that lim cλ = ∞

(5.50)

λ↑σ1

since lim θ[λ,Ω] (xλ ) = b.

λ↑σ1

By the definition of cλ , for each λ ∈ (σˆ 1 , σ1 ), we have that

2 2 θ[λ,Ω] (x)
θ[λ,Ω] (xλ ) + cλ e−αR /4 − e−αR

 R/2 Y (xλ ) . ∀x ∈ B

(5.51)

Now, for each λ ∈ (σˆ 1 , σ1 ), we consider the auxiliary function vλ := θ[λ,Ω] − θ[λ,Ω] (xλ ) − cλ ψλ Thanks to (5.51),



 R/2 Y (xλ ) . in AR = BR Y (xλ ) \ B



vλ
0 on ∂BR/2 Y (xλ ) .

Moreover, since ψλ = 0 on ∂BR (Y (xλ )), it follows from (5.48) that

 vλ = θ[λ,Ω] − θ[λ,Ω] (xλ )
0 on ∂BR Y (xλ ) .

Furthermore, due to (5.49), in AR we have that

(−
− λ)vλ = λθ[λ,Ω] (xλ ) − cλ (−
− λ)ψλ
λθ[λ,Ω] (xλ ) + cλ ω and hence, by (5.50),



 R/2 Y (xλ ) (−
− λ)vλ > 0 in AR = BR Y (xλ ) \ B

provided λ < σ1 is sufficiently close to σ1 . Therefore, since λ < σ1 = σ [−
, Ω0,1 ] < σ [−
, AR ], it follows from Theorem 3.1 that vλ (x) > 0 if and consequently,

 R  < x − Y (xλ ) < R, 2

θ[λ,Ω] (x)
θ[λ,Ω] (xλ ) + cλ ψλ (x) for each x ∈ AR .

(5.52)

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

269

Now, setting Y (xλ ) − xλ , R

nλ := we have that

∂θ[λ,Ω] θ[λ,Ω] (xλ + tnλ ) − θ[λ,Ω] (xλ ) (xλ ) = lim . t↓0 ∂nλ t Moreover, thanks to (5.52), for each t ∈ (0, R2 ), we find that θ[λ,Ω] (xλ + tnλ ) − θ[λ,Ω] (xλ ) t cλ ψλ (xλ + tnλ )
t 2

2

=

cλ (e−α|xλ +tnλ −Y (xλ )| − e−αR ) t

=

cλ (e−α|tnλ −Rnλ | − e−αR ) t

=

cλ (e−α(R−t) − e−αR ) , t

2

2

2

2

and hence, since 2

lim t↓0

2

e−α(R−t) − e−αR 2 = 2αRe−αR , t

we obtain that ∂θ[λ,Ω] 2 (xλ )
2αRe−αR cλ . ∂nλ Therefore, thanks to (5.50), lim

λ↑σ1

∂θ[λ,Ω] (xλ ) = ∞. ∂nλ

(5.53)

Subsequently, for each λ < σ1 , λ ∼ σ1 , we consider the auxiliary boundary value problem


−
u = λu − af (·, u)u u = θ[λ,Ω] (xλ )

in D, on ∂D.

(5.54)

270

J. López-Gómez

Since λ < σ2 , thanks to Lemma 5.1, (5.54) possesses a unique positive solution, denoted by ϑλ := θ[λ,D,θ[λ,Ω] (xλ )] . Moreover, since θ[λ,Ω] |D provides us with a positive supersolution of (5.54), we have that ϑλ  θ[λ,Ω]

 in D.

Therefore, since ϑλ (xλ ) = θ[λ,Ω] (xλ ), we find that ∂ϑλ ∂θ[λ,Ω] (xλ )
(xλ ), ∂nλ ∂nλ and consequently, by (5.53), lim

λ↑σ1

∂ϑλ (xλ ) = ∞. ∂nλ

(5.55)

 as λ ↑ σ1 , the unique This is impossible, since the functions ϑλ approximate in C 1 (D), positive solution, ϑσ1 := θ[σ1 ,D,b] , of


−
u = σ1 u − af (·, u)u u=b

in D, on ∂D.

Therefore, (5.45) get shown, which completes the proof of Theorem 2.3. Finally, we will complete the proof of Theorem 2.2. It remains to prove (2.5). Thanks to (5.37), it suffices to prove that lim min Lmin [λ,D] = ∞,

λ↑σ2 ∂Ω0,2

(5.56)

0,1 , and that, for some large solution L of (1.9) in Ω− , where D = Ω \ Ω lim Lmin [λ,D] = L in Ω− .

λ↑σ2

(5.57)

We already know that, for each λ < σ2 , Lmin [λ,D]
θ[λ,D] . Moreover, the proof of (5.45) can be easily adapted to show that lim min θ[λ,D] = ∞;

λ↑σ2 ∂Ω0,2

now, one must work with ∂Ω0,1 , instead of ∂D, though the technical details can be adapted mutatis mutandis. Therefore (5.56) holds true.

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

271

In order to prove (5.57), we consider the point-wise limit (5.42), which is well defined 0,2 , and, for sufficiently large n ∈ N, the open sets and equals ∞ in Ω  n := x ∈ Ω− : dist(x, ∂Ω− ) > n−1 . Ω−

Subsequently, for each λ < σ2 , we set Mλ := max Lmin [λ,D] . n ∂Ω−

n By definition, Lmin [λ,D] |∂Ω−  Mλ . Hence, thanks to Lemma 4.10,

min n ,M ]  θ[σ ,Ω n ,M ]  L Lmin [λ,D]  θ[λ,Ω− λ λ 2 [σ2 ,Ω n ] − −

n in Ω− .

Therefore, due to Theorem 4.11, we obtain that, for each λ < σ2 , max Lmin [λ,D]  L[σ2 ,Ω− ]

in Ω− .

(5.58)

Consequently, the point-wise limit (5.42) is finite in Ω− . Moreover, by the monotonicity in λ and the compactness of (−
)−1 , it provides us with a large solution of (1.9) in Ω− , which concludes the proof of Theorem 2.2.

6. Proof of Theorem 2.4 Part (a) is a direct consequence from the last assertion of Theorem 3.5, as well as part (b) since, due to Theorem 2.3, (1.1) possesses a positive steady state for each λ ∈ (σ0 , σ1 ). So, it remains to prove parts (c) and (d). Suppose σ1  λ < σ2 . Thanks to the parabolic maximum principle, for each ε > 0 and t
0, u[λ,Ω] (·, t; u0 )
u[σ1 −ε,Ω] (·, t; u0 ) since λ > σ1 − ε, and hence, thanks to part (b), lim inf u[λ,Ω] (·, t; u0 )
lim u[σ1 −ε,Ω] (·, t; u0 ) = θ[σ1 −ε,Ω] . t↑∞

t↑∞

(6.1)

As (6.1) holds true for each sufficiently small ε > 0, it follows from Theorem 2.3 that lim inf u[λ,Ω] (·, t; u0 )
lim θ[σ1 −ε,Ω] = Mmin  [σ ,Ω\Ω t↑∞

ε↓0

1

0,1 ]

.

(6.2)

272

J. López-Gómez

In particular, lim inf u[λ,Ω] (·, t; u0 ) = ∞ t↑∞

0,1 \ ∂Ω, uniformly in compact subsets of Ω

(6.3)

and hence, for each M > 0 there exists a constant TM > 0 such that u[λ,Ω] (x, t; u0 )
M

for each (x, t) ∈ Γ1 × [TM , ∞).

Thus, u[λ,Ω] (·, t; u0 ) provides us with a supersolution of the parabolic problem  ∂u

 0,1 × (TM , ∞), in Ω \ Ω   ∂t −
u = λu − af (·, u)u u=M on Γ1 × (TM , ∞),   0,1 . u(·, TM ) = u[λ,Ω] (·, TM ; u0 ) in Ω \ Ω

By the parabolic maximum principle, we have that

 u[λ,Ω] (x, t; u0 )
u[λ,Ω\Ω0,1 ,M] x, t − TM ; u[λ,Ω] (·, TM ; u0 )

0,1 × (TM , ∞). So, thanks to Lemma 5.1, for each (x, t) ∈ Ω \ Ω

 lim inf u[λ,Ω] (x, t; u0 )
lim u[λ,Ω\Ω0,1 ,M] x, t − TM ; u[λ,Ω] (·, TM ; u0 ) t↑∞

t↑∞

= θ[λ,Ω\Ω0,1 ,M] .

Therefore, passing to the limit as M ↑ ∞, lim inf u[λ,Ω] (x, t; u0 )
Lmin  [λ,Ω\Ω

0,1 ]

t↑∞

0,1 in Ω \ Ω

since Lmin  [λ,Ω\Ω

0,1 ]

= lim θ[λ,Ω\Ω0,1 ,M] M↑∞

(see the proof of Theorem 2.2). Consequently, bringing together this estimate with (6.3) we find that lim inf u[λ,Ω] (x, t; u0 )
Mmin  [λ,Ω\Ω

0,1 ]

t↑∞

in Ω.

(6.4)

This estimate provides us with Theorem 2.4(c)(i) and the lower estimate of Theorem 2.4(c)(ii). Now, we assume, in addition, that u0 is a subsolution of (1.9) in Ω. Then, for each t > 0, the function x → u[λ,Ω] (x, t; u0 ),

 x ∈ Ω,

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

273

is a subsolution of (1.9) in Ω, since t → u[λ,Ω] (·, t; u0 ) is increasing. Fix t > 0 and set 0,1 . D := Ω \ Ω

Mt := max u[λ,Ω] (·, t; u0 ),  D

Thanks to Lemma 5.1, for each M
Mt , we have that 0,1 . in Ω \ Ω

u[λ,Ω] (·, t; u0 )  θ[λ,Ω\Ω0,1 ,M]

Hence, by the construction of the minimal large solution, u[λ,Ω] (·, t; u0 )  lim θ[λ,Ω\Ω0,1 ,M] = Lmin  [λ,Ω\Ω

0,1 ]

M↑∞

0,1 , in Ω \ Ω

and therefore, passing to the limit as t ↑ ∞, lim sup u[λ,Ω] (·, t; u0 )  Lmin  [λ,Ω\Ω

0,1 ]

t↑∞

0,1 . in Ω \ Ω

Consequently, due to (6.4), lim u[λ,Ω] (·, t; u0 ) = Mmin  [λ,Ω\Ω

0,1 ]

t↑∞

in Ω,

(6.5)

which concludes the proof of Theorem 2.4(c)(iii). Now, suppose u0 is arbitrary – not necessarily a subsolution of (1.9) in Ω – and λ
σ2 . Then, thanks to the parabolic maximum principle, we have that u[λ,Ω] (·, t; u0 )
u[σ2 −ε,Ω] (·, t; u0 )

in Ω,

and hence, it follows from (6.4) that lim inf u[λ,Ω] (·, t; u0 )
lim inf u[σ2 −ε,Ω] (·, t; u0 )
Mmin  [σ −ε,Ω\Ω t↑∞

t↑∞

2

0,1 ]

(6.6)

in Ω for each sufficiently small ε > 0 since σ1 < σ 2 − ε < σ 2 . Note that we cannot apply (6.5) since u0 might not be a subsolution of (1.9) in Ω. As (6.6) is satisfied for every sufficiently small ε > 0, and, thanks to Theorem 2.2, there exists a metasolution of −
u = σ2 u − af (·, u)u

274

J. López-Gómez

supported in Ω− , say M[σ2 ,Ω− ] , such that lim Mmin  [σ −ε,Ω\Ω ε↓0

0,1 ]

2

= M[σ2 ,Ω− ] ,

it follows from (6.6) that lim inf u[λ,Ω] (·, t; u0 )
M[σ2 ,Ω− ]

in Ω.

t↑∞

In particular, lim inf u[λ,Ω] (·, t; u0 ) = ∞ in Ω \ Ω− t↑∞

(6.7)

which concludes the proof of Theorem 2.4(d)(i). Moreover, since (6.7) occurs uniformly on ∂Ω− , for each M > 0, there exists a time TM > 0 for which u[λ,Ω] (x, t; u0 )
M

if (x, t) ∈ ∂Ω− × (TM , ∞).

Thus, u[λ,Ω] (·, t; u0 ) provides us with a supersolution of the parabolic problem  ∂u in Ω− × (TM , ∞),   ∂t −
u = λu − auf (·, u) u=M on ∂Ω− × (TM , ∞),   u(·, TM ) = u[λ,Ω] (·, TM ; u0 ) in Ω− .

By the parabolic maximum principle, we have that, for each (x, t) ∈ Ω− × (TM , ∞),

 u[λ,Ω] (x, t; u0 )
u[λ,Ω− ,M] x, t − TM ; u[λ,Ω] (·, TM ; u0 ) , and hence, thanks to Theorem 4.9,

 lim inf u[λ,Ω] (x, t; u0 )
lim u[λ,Ω− ,M] x, t − TM ; u[λ,Ω] (·, TM ; u0 ) t↑∞

t↑∞

= θ[λ,Ω− ,M] .

Therefore, passing to the limit as M ↑ ∞ gives lim inf u[λ,Ω] (·, t; u0 )
Lmin [λ,Ω− ] t↑∞

in Ω−

since Lmin [λ,Ω− ] = lim θ[λ,Ω− ,M] . M↑∞

Consequently, bringing together this estimate with (6.7) gives lim inf u[λ,Ω] (·, t; u0 )
Mmin [λ,Ω− ] t↑∞

in Ω

(6.8)

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

275

which provides us with the lower estimate of Theorem 2.4(d)(ii). Now, besides λ
σ2 , suppose u0 is a subsolution of (1.9) in Ω. Then, for each t > 0, the function u[λ,Ω] (·, t; u0 ) is a subsolution of (1.9) in Ω, since t → u[λ,Ω] (·, t; u0 ) is increasing. Fix t > 0 and set Mt := max u[λ,Ω] (·, t; u0 ). − Ω

Thanks to Lemma 4.10, for each M
Mt , we have that u[λ,Ω] (·, t; u0 )  θ[λ,Ω− ,M]

in Ω− ,

and hence, u[λ,Ω] (·, t; u0 )  lim θ[λ,Ω− ,M] = Lmin [λ,Ω− ] M↑∞

in Ω− .

Therefore, passing to the limit as t ↑ ∞, lim sup u[λ,Ω] (·, t; u0 )  Lmin [λ,Ω− ] t↑∞

in Ω−

and, consequently, due to (6.7) and (6.8), lim u[λ,Ω] (·, t; u0 ) = Mmin [λ,Ω− ]

t↑∞

in Ω,

which concludes the proof of Theorem 2.4(d)(iii). To conclude the proof of Theorem 2.4, it remains to obtain the upper estimates for an arbitrary u0 > 0, in order to get the upper estimates of parts (c)(ii) and (d)(ii). To do this, the strategy adopted here consists in obtaining a priori bounds in Ω− for the solutions of (1.1). These a priori bounds can be obtained as follows. Fix λ
σ1 and consider the function u[λ,Ω] (·, 1; u0 ) ≫ 0. Then, there exists κ > 1 such that u[λ,Ω] (·, 1; u0 ) < κϕ,

(6.9)

where ϕ ≫ 0 is a principal eigenfunction associated with σ0 . We claim that there exists Λ > max{λ, σ2 }

(6.10)

for which the function κϕ is a subsolution of
−
u = Λu − af (·, u)u in Ω, u=0 on ∂Ω. Indeed, since κϕ = 0 on ∂Ω, κϕ is a subsolution of (6.11) if and only if −
(κϕ)  Λκϕ − af (·, κϕ)κϕ

in Ω,

(6.11)

276

J. López-Gómez

or equivalently, af (·, κϕ)  Λ − σ0

in Ω

which is satisfied for any sufficiently large Λ satisfying (6.10). Now, thanks to the parabolic maximum principle, it follows from (6.9) that, for any (x, t) ∈ Ω × (0, ∞),

 u[λ,Ω] (x, t + 1; u0 ) = u[λ,Ω] x, t; u[λ,Ω] (·, 1; u0 )  u[λ,Ω] (·, t; κϕ).

Similarly, (6.10) implies

u[λ,Ω] (·, t; κϕ) < u[Λ,Ω] (·, t; κϕ). Thus, for each t > 0, u[λ,Ω] (·, t + 1; u0 )  u[Λ;Ω] (·, t; κϕ) in Ω.

(6.12)

As κϕ is a subsolution of (6.11), it follows from part (d)(iii) that lim u[Λ,Ω] (·, t; κϕ) = Lmin [Λ,Ω− ]

t↑∞

in Ω− ,

and hence, (6.12) implies lim sup u[λ,Ω] (·, t; u0 )  Lmin [Λ,Ω− ] t↑∞

in Ω− .

(6.13)

Consequently, u[λ,Ω] (·, t; u0 ) is uniformly bounded above in any compact subset of Ω− for each t > 0, which provides us with the necessary a priori bounds to complete the proof of the theorem. Subsequently, we suppose σ1  λ < σ2 , and for each sufficiently large n ∈ N, we consider the open set


1  . Dn := x ∈ D := Ω \ Ω0,1 : dist(x, ∂D) > n Fix one of these values of n. Since ∂Dn ⊂ Ω− , it follows from (6.13) that there exists a constant M0 > 0 such that, for each M
M0 and t > 0, u[λ,Ω] (·, t; u0 )  M

on ∂Dn ,

and hence, the parabolic maximum principle shows that, for each t > 0, u[λ,Ω] (·, t; u0 )  u[λ,Dn ,M] (·, t; u0 )

in Dn ,

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

277

where u[λ,Dn ,M] (·, t; u0 ) is the solution defined in Lemma 5.2. Due to Lemma 5.2, we have that lim u[λ,Dn ,M] (·, t; u0 ) = θ[λ,Dn ,M]

in Dn

t↑∞

and therefore, lim sup u[λ,Ω] (·, t; u0 )  θ[λ,Dn ,M]

in Dn .

(6.14)

t↑∞

Consequently, passing to the limit as M ↑ ∞ in (6.14) gives lim sup u[λ,Ω] (·, t; u0 )  Lmin [λ,Dn ]

in Dn .

(6.15)

t↑∞

As (6.15) is valid for all sufficiently large n ∈ N, and, thanks to the analysis carried out in the proof of Theorem 2.2, we already know that min Lmax [λ,D] = lim L[λ,Dn ] , n↑∞

we find from (6.15) that lim sup u[λ,Ω] (·, t; u0 )  Lmax [λ,D]

in D,

t↑∞

which concludes the proof of part (c)(ii). Finally, suppose λ
σ2 , and, for each sufficiently large n ∈ N, consider the open subset of Ω− defined by n Ω−


1 := x ∈ Ω− : dist(x, ∂Ω− ) > n

n ⊂ Ω , it follows from (6.13) that there exists and fix one of these values of n. Since ∂Ω− − a constant M0 > 0 such that, for each M
M0 and t > 0,

u[λ,Ω] (·, t; u0 )  M

n on ∂Ω− .

Thus, by the parabolic maximum principle, we have that, for each t > 0, u[λ,Ω] (·, t; u0 )  u[λ,Ω−n ,M] (·, t; u0 )

n in Ω− ,

278

J. López-Gómez

where u[λ,Ω−n ,M] (·, t; u0 ) is the unique solution of the parabolic problem  ∂u   ∂t −
u = λu − af (·, u)u u=M   u(·, 0) = u0

n × (0, ∞), in Ω−

n × (0, ∞), on ∂Ω− n in Ω− .

Thanks to Theorem 4.9,

lim u[λ,Ω−n ,M] (·, t; u0 ) = θ[λ,Ω−n ,M]

t↑∞

n in Ω−

and therefore, lim sup u[λ,Ω] (·, t; u0 )  θ[λ,Ω−n ,M] t↑∞

n in Ω− .

(6.16)

Consequently, passing to the limit as M ↑ ∞ in (6.16) gives lim sup u[λ,Ω] (·, t; u0 )  Lmin [λ,Ω n ] −

t↑∞

n in Ω− .

(6.17)

As (6.17) is valid for all sufficiently large n ∈ N and we already know that min Lmax [λ,Ω− ] = lim L[λ,Ω n ] , n↑∞

−

(6.17) implies lim sup u[λ,Ω] (·, t; u0 )  Lmax [λ,Ω− ] t↑∞

in Ω−

which concludes the proof of part (d)(ii). The proof of the theorem is completed.

7. Proofs of Theorems 2.6–2.8 The proofs of Theorems 2.6 and 2.7 are based upon the following estimates for the boundary growth rate of the large positive solutions of −
u = λu − aup

(7.1)

0,1 , Ω− }. in D ∈ {Ω \ Ω P ROPOSITION 7.1. Under the assumptions of Theorem 2.6, for each λ < σ2 and any large 0,1 , one has that positive solution L(x) of (7.1) in Ω \ Ω lim t↓0

L(x1 − tnx1 ) =1 B(x1 )[dist(x1 − tnx1 , Γ1 )]−α(x1 )

for each x1 ∈ Γ1 ,

(7.2)

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

279

0,1 at x1 and where nx1 stands for the outward unit normal to Ω \ Ω α(x1 ) :=

γ (x1 ) + 2 , p−1

B(x1 ) :=

'

α(x1 )(α(x1 ) + 1) β(x1 )

(1/(p−1)

.

(7.3)

0,1 , one has Therefore, for any pair (L1 , L2 ) of large positive solutions of (7.1) in Ω \ Ω that lim t↓0

L1 (x1 − tnx1 ) = 1 for each x1 ∈ Γ1 . L2 (x1 − tnx1 )

(7.4)

Notice that, for each x1 ∈ Γ1 and sufficiently small t > 0, dist(x1 − tnx1 , Γ1 ) = t. P ROPOSITION 7.2. Under the assumptions of Theorem 2.7, for each λ ∈ R and any large positive solution L(x) of (7.1) in Ω− , one has that lim t↓0

L(x1 − tnx1 ) = 1 for each x1 ∈ ∂Ω− , B(x1 )[dist(x1 − tnx1 , ∂Ω− )]−α(x1 )

(7.5)

where nx1 stands for the outward unit normal to Ω− at x1 and α(x1 ) and B(x1 ) are given through (7.3). Therefore, for any pair (L1 , L2 ) of large positive solutions of (7.1) in Ω− , one has that lim t↓0

L1 (x1 − tnx1 ) = 1 for each x1 ∈ ∂Ω− . L2 (x1 − tnx1 )

(7.6)

The distribution of this section is as follows. In Section 7.1 we show that in one spatial dimension the estimates (7.2) and (7.5) follow in a rather natural manner. In Section 7.2 we characterize the boundary blow-up rates of the large solutions of (7.1) for some related radially symmetric problems. In Section 7.3, we prove Propositions 7.1 and 7.2. In Section 7.4 we complete the proofs of Theorems 2.6 and 2.7. Finally, in Section 7.5 we prove Theorem 2.8.

7.1. Finding out the boundary blow-up rate of a large solution In one space dimension, N = 1, as well as in the radially symmetric case, the fact that (2.9) and (2.10) imply (7.2) follows in a very natural manner. Suppose that we want ascertaining the blow up rate at R > 0 of any solution u(x) of the one-dimensional singular boundary value problem


−u′′ = λu − aup in (0, R), u(0) = 0, limx↑R u(x) = ∞,

(7.7)

280

J. López-Gómez

where, for some γ
0, a(x) = β(x)(R − x)γ ,

x ∈ (0, R), β(R) = 0.

Then, the change of variable u(x) = (R − x)−α ψ(x),

x ∈ [0, R],

where α > 0 has to be determined, transforms (7.7) into the differential equation (R − x)−α ψ ′′ (x) + 2α(R − x)−α−1 ψ ′ (x) + α(α + 1)(R − x)−α−2 ψ(x) = β(x)(R − x)γ −αp ψ p (x) − λ(R − x)−α ψ(x)

(7.8)

subject to the boundary conditions ψ(0) = 0,

ψ(R) ∈ (0, ∞),

so that α provides us with the exact blow-up rate of u at R. Multiplication by (R − x)α+2 transforms (7.8) into (R − x)2 ψ ′′ (x) + 2α(R − x)ψ ′ (x) + α(α + 1)ψ(x)

= β(x)(R − x)γ −αp+α+2 ψ p (x) − λ(R − x)2 ψ(x).

Thus, assuming % % & & lim (R − x)2 ψ ′′ (x) = lim (R − x)ψ ′ (x) = 0,

x↑R

x↑R

one is driven to impose γ − αp + α + 2 = 0,

α(α + 1)ψ(R) = β(R)ψ p (R),

which provides us with the values of α and ψ(R). Namely,

α :=

γ +2 p−1

and ψ(R) =

'

α(α + 1) β(R)

(1/(p−1)

in complete agreement with the statement of Propositions 7.1 and 7.2.

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

281

7.2. Two auxiliary radially symmetric problems The main result of this section is the following lemma. L EMMA 7.3. Suppose x0 ∈ RN , R > 0, λ ∈ R, p > 1, γ
0 and β(x) = β(r),

 r := |x − x0 |, β ∈ C [0, R]; (0, ∞) .

Then, for each ε > 0, the singular boundary value problem


% &γ −
u = λu − β(x) dist x, ∂BR (x0 ) up u=∞

in BR (x0 ), on ∂BR (x0 )

(7.9)

possesses a positive solution Lε such that 1 − ε  lim inf d(x)↓0

Lε (x) Lε (x) −α  lim sup −α  1 + ε, B[d(x)] B[d(x)] d(x)↓0

(7.10)

where

and

 d(x) := dist x, ∂BR (x0 ) = R − |x − x0 | = R − r α :=

γ +2 , p−1

B :=

'

α(α + 1) β(R)

(1/(p−1)

.

(7.11)

P ROOF. Note that the radially symmetric solutions of (7.9) are given by u(x) := ψ(r), where ψ satisfies  −ψ ′′ −

r = |x − x0 |,

N −1 ′ r ψ

ψ ′ (0) = 0,

= λψ − β(r)(R − r)γ ψ p

in (0, R),

limr↑R ψ(r) = ∞.

(7.12)

First, we show that, for each sufficiently small ε > 0, there exists a constant Aε > 0 such that, for every A > Aε , the function ψ¯ ε (r) := A + B+

 2 r (R − r)−α , R

B+ := (1 + ε)B,

provides us with a positive supersolution of (7.12). Indeed, ψ¯ ε′ (0) = 0 and

lim ψ¯ ε (r) = ∞

r↑R

(7.13)

282

J. López-Gómez

since α > 0. Thus, ψ¯ ε is a supersolution of (7.12) if and only if B+ B+ (R − r)−α − α(N + 3) 2 r(R − r)−α−1 R2 R  2 r (R − r)−α−2 − α(α + 1)B+ R  2 ( ' r α −α A(R − r) + B+
λ(R − r) R  2 (p ' r α γ −αp A(R − r) + B+ − β(r)(R − r) . R

−2N

Thus, multiplying this inequality by (R − r)α+2 and taking into account that α + 2 + γ − αp = 0, we find that ψ¯ ε is a supersolution of (7.12) if and only if  2 B+ r B+ 2 (R − r) − α(N + 3) r(R − r) − α(α + 1)B + R R2 R2  2 (  2 (p ' ' r r α α 2 − β(r) A(R − r) + B+
λ(R − r) A(R − r) + B+ . R R (7.14)

−2N

At r = R, (7.14) becomes into p

−α(α + 1)B+
−β(R)B+ , which is satisfied if and only if p−1

B+




α(α + 1) . β(R)

Therefore, by making the choice (7.13), the inequality (7.14) is satisfied in a left neighborhood of r = R, say (R − δ, R], for some δ = δ(ε) > 0. Finally, by choosing a sufficiently large A, it is clear that the inequality is satisfied in the whole interval [0, R] since p > 1 and β is bounded away from zero. This concludes the proof of the claim above. Now, we will construct an appropriate subsolution for problem (7.12). For doing this we shall distinguish two different cases according to the sign of the parameter λ. First, we assume λ
0.

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

283

Then, for each sufficiently small ε > 0, there exists C < 0 for which the function

 2
r ψ ε (r) := max 0, C + B− (R − r)−α R provides us with a nonnegative subsolution of (7.12) if B− = (1 − ε)B.

(7.15)

Indeed, it is easy to see that ψ ε is a subsolution of (7.12) if in the region where  2 r C + B− (R − r)−α
0, R the following inequality is satisfied B− B− (R − r)−α − α(N + 3) 2 r(R − r)−α−1 R2 R  2 r − α(α + 1)B− (R − r)−α−2 R '  2 ( r  λ(R − r)−α C(R − r)α + B− R '  2 (p r γ −αp α C(R − r) + B− . − β(r)(R − r) R

−2N

Equivalently,  2 r B− B− 2 (R − r) − α(N + 3) r(R − r) − α(α + 1)B − 2 2 R R R  2 (  2 (p ' ' r r − β(r) C(R − r)α + B−  λ(R − r)2 C(R − r)α + B− . R R (7.16)

−2N

Now, note that for each C < 0 there exists a constant z = z(C) ∈ (0, R) such that  2 %  r (R − r)−α < 0 if r ∈ 0, z(C) C + B− R

284

J. López-Gómez

and  2 %  r C + B− (R − r)−α
0 if r ∈ z(C), R R because the mapping  2 r (R − r)−α r → R is increasing. Moreover, z(C) is decreasing with respect to C and lim z(C) = R,

C↓−∞

lim z(C) = 0.

(7.17)

C↑0

Thus, since λ
0, for each r ∈ [z(C), R), the following condition implies (7.16)  2 (p  2 ' r r β(r) C(R − r)α + B−  α(α + 1)B− . R R

(7.18)

Note that, since C < 0, for (7.18) to be satisfied it suffices that p−1 β(r)B−

 2(p−1) r  α(α + 1) R

(7.19)

for each r ∈ [z(C), R). Making the choice (7.15) and using the continuity of β(r), it is easy to see that there exists a constant δ(ε) > 0 for which (7.19) is satisfied in [R − δ(ε), R). Moreover, thanks to (7.17), there exists C < 0 such that z(C) = R − δ(ε). For this choice of C, it readily follows that ψ ε provides us with a subsolution of (7.12). Similarly, it is easy to see that there exists C < 0 for which the function ψ ε (r) := e

√ −λ(r−R)

 2


r −α max 0, C + B− (R − r) R

provides us with a subsolution of (7.12) if λ < 0. Note that in each of these cases ψ¯ ε (r) −α = 1, r↑R B+ (R − r) lim

lim

r↑R

ψ ε (r) −α

B− (R − r)

= 1,

where B+ and B− are the constants defined in (7.13) and (7.15). Moreover, for any sufficiently large A > Aε , ψ ε  ψ¯ ε ,

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

285

and hence, 1 − ε = lim r↑R

ψ ε (r) −α

B(R − r)

ψ¯ ε (r) −α = 1 + ε. r↑R B(R − r)

 lim

Therefore, setting L ε (x) := ψ ε (r),

ε (x) := ψ¯ ε (r), L

r := |x − x0 |,

ε ) provides us with an ordered sub–supersolution pair of (7.9) such that ( L ε, L ε (x) L L ε (x) −α  lim −α = 1 + ε. d(x)↓0 B[d(x)] d(x)↓0 B[d(x)]

1 − ε = lim

(7.20)

Consequently, by Theorem 3.3, there exists a positive solution Lε of (7.9) such that ε . L ε  Lε  L Moreover, due to (7.20), Lε must satisfy (7.10), which concludes the proof. Now, adapting the previous argument together with a reflection around r0 := Lemma 7.3 provides us with the corresponding result for each of the annuli

 R1 +R2 2 ,

 AR1 ,R2 (x0 ) := x ∈ RN : 0 < R1 < |x − x0 | < R2 .

L EMMA 7.4. Suppose x0 ∈ RN , R2 > R1 > 0, λ ∈ R, p > 1, γ
0 and β(x) = β(r),

 r := |x − x0 |, β ∈ C [R1 , R2 ]; (0, ∞) ,

2 of some function β˜ ∈ C([r0 , R2 ]; (0, ∞)), so that, in is the reflection around r0 := R1 +R 2 particular, β(R1 ) = β(R2 ). Then, for each ε > 0, the singular problem,



% &γ −
u = λu − β(x) dist x, ∂AR1 ,R2 (x0 ) up

u=∞

in AR1 ,R2 (x0 ), on ∂AR1 ,R2 (x0 ),

(7.21)

possesses a positive solution Lε (x) satisfying 1 − ε  lim inf δ(x)↓0

Lε (x) Lε (x) −α  lim sup −α  1 + ε, B[δ(x)] δ(x)↓0 B[δ(x)]

where α and B are given by (7.11) and


 R2 − |x − x0 | δ(x) := dist x, ∂AR1 ,R2 (x0 ) = |x − x0 | − R1

if r0  |x − x0 | < R2 ,

if R1 < |x − x0 | < r0 .

(7.22)

286

J. López-Gómez

7.3. Proof of Propositions 7.1 and 7.2 As it will be clear later, the proof of Proposition 7.1 can be easily adapted to prove Proposition 7.2, as it is of a local nature around each of the points of the boundary of the underlying domain. So, suppose (2.9), (2.10), λ < σ2 and L is a large positive solution of (7.1) 0,1 . Pick in Ω \ Ω

 0,1 . x1 ∈ Γ1 = ∂ Ω \ Ω

Since Γ1 is of class C 2 , there exist R > 0 and δ0 > 0 such that R (x1 − Rnx1 ) ∩ Γ1 = {x1 }, B

BR (x1 − Rnx1 ) ⊂ Ω− ,

(7.23)

and, for each δ ∈ (0, δ0 ],



R x1 − (R + δ)nx1 ⊂ Ω− . B

(7.24)

On the other hand, due to (2.10), given any η ∈ (0, β(x1 )), R > 0 can be shortened, if necessary, so that, for each δ ∈ [0, δ0 ],

% &γ (x1 ) a
β(x1 ) − η dist(·, Γ1 )



in BR x1 − (R + δ)nx1 .

(7.25)

As for each δ ∈ [0, δ0 ] and x ∈ BR (x1 − (R + δ)nx1 ), we have that 

dist(x, Γ1 )
dist x, BR x1 − (R + δ)nx1  % & = R − x − x1 − (R + δ)nx1 ,

(7.25) implies

where

 γ (x1 ) a(x)
β(x1 ) − η R − rδ (x) ,



x ∈ BR x1 − (R + δ)nx1 ,

(7.26)

 % & rδ (x) := x − x1 − (R + δ)nx1 .

Thanks to (7.24) and (7.26), for each δ ∈ (0, δ0 ], the restriction Lδ := L|BR (x1 −(R+δ)nx

1

(7.27)

)

provides us with a classical subsolution (bounded) of the singular problem 

 −
u = λu − β(x1 ) − η (R − rδ )γ (x1 ) up u=∞



in BR x1 − (R + δ)nx1 , 

on ∂BR x1 − (R + δ)nx1 .

(7.28)

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

287

Now, consider the limiting problem at δ = 0


 −
u = λu − β(x1 ) − η (R − r0 )γ (x1 ) up u=∞

in BR (x1 − Rnx1 ), on ∂BR (x1 − Rnx1 ).

(7.29)

Subsequently, for each δ ∈ [0, δ0 ], we shall denote



dδ (x) := dist x, ∂BR x1 − (R + δ)nx1 

 = R − x − x1 − (R + δ)nx1  = R − rδ (x).

Thanks to Lemma 7.3, for each ε > 0, the problem (7.29) possesses a solution, say Lε , such that lim sup d0 (x)↓0

Lε (x) −α(x1 )

Bη (x1 )[d0 (x)]

 1 + ε,

(7.30)

where α(x1 ) is given by (7.3) and Bη (x1 ) :=

'

α(x1 )(α(x1 ) + 1) β(x1 ) − η

(1/(p−1)

.

Subsequently, we fix ε > 0 and, for each δ ∈ (0, δ0 ], consider the function Lδε defined by Lδε (x) := Lε (x + δnx1 ), By construction,



x ∈ BR x1 − (R + δ)nx1 .

lim Lδε = Lε δ↓0

(7.31)

and, for each δ ∈ (0, δ0 ], Lδε provides us with a positive solution of (7.28). Thus, since the function Lδ defined in (7.27) is a positive classical subsolution of (7.28), it follows from Lemma 3.6 that, for each δ ∈ (0, δ0 ], Lδ  Lδε



in BR x1 − (R + δ)nx1 .

Consequently, passing to the limit as δ ↓ 0, (7.27) and (7.31) imply L  Lε

in BR (x1 − Rnx1 ).

In particular, for each t ∈ (0, R), L(x1 − tnx1 )  Lε (x1 − tnx1 )

288

J. López-Gómez

and hence, L(x1 − tnx1 ) Lε (x1 − tnx1 )  . −α(x ) 1 Bη (x1 )[dist(x1 − tnx1 , Γ1 )] Bη (x1 )[dist(x1 − tnx1 , Γ1 )]−α(x1 ) On the other hand, by construction, we have that dist(x1 − tnx1 , Γ1 ) = t = d0 (x1 − tnx1 ) and so, passing to the limit as t ↓ 0 in the previous inequality, (7.30) gives lim sup t↓0

L(x1 − tnx1 )  1 + ε. Bη (x1 )[dist(x1 − tnx1 , Γ1 )]−α(x1 )

As this inequality holds for each sufficiently small η > 0 and ε > 0, it is apparent that lim sup t↓0

L(x1 − tnx1 )  1, B(x1 )[dist(x1 − tnx1 , Γ1 )]−α(x1 )

(7.32)

where α and B are given by (7.3). Consequently, to complete the proof of (7.2) it remains to show that 1  lim inf t↓0

L(x1 − tnx1 ) . B(x1 )[dist(x1 − tnx1 , Γ1 )]−α(x1 )

(7.33)

To prove (7.33) we will construct a large subsolution having the appropriate growth at x1 ∈ Γ1 . Since Γ1 is smooth, there exist R2 > R1 > 0 and δ0 > 0 such that R1 ,R2 (x1 + R1 nx1 ) ∩ Γ1 = {x1 }, A , 

0,1 ⊂ AR1 ,R2 x1 + (R1 + δ)nx1 Ω \Ω δ∈[0,δ0 ]

and 0,1 ⊂ AR1 ,(R1 +R2 )/2 (x1 + Rnx1 ), Ω \Ω which can be accomplished by taking a sufficiently small R1 > 0 and a sufficiently large R2 > R1 . Now, fix a sufficiently small η > 0. Thanks to (2.10), there exists a radially symmetric function R1 ,R2 (x1 + R1 nx1 ) → [0, ∞) a˜ : A such that a  a˜  sup a + 1 0,1 Ω\Ω

0,1 , in Ω \ Ω

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

289

and, for each x ∈ AR1 ,R2 (x1 + R1 nx1 ),

% &γ (x1 ) a(x) ˜ = b |x − x1 − R1 nx1 | dist x, ∂AR1 ,R2 (x1 + R1 nx1 )

for some b ∈ C([R1 , R2 ]; (0, ∞)) satisfying b(R1 ) = β(x1 ) + η.

Moreover, by enlarging R2 , if necessary, one can assume that b is the reflection around R1 +R2 of a continuous positive function. Actually, by shortening δ0 , if necessary, we can 2 assume that it satisfies all the requirements of the proof of (7.32). Subsequently, for each δ ∈ [0, δ0 ], we consider the singular auxiliary problem 



in AR1 ,R2 x1 + (R1 + δ)nx1 , 

on ∂AR1 ,R2 x1 + (R + δ)nx1 .

−
u = λu − a(· ˜ + δnx1 )up u=∞

(7.34)

Fix a sufficiently small ε > 0. Thanks to Lemma 7.4, for δ = 0, (7.34) possesses a positive solution Lε such that 1 − ε  lim inf

D0 (x)↓0

Lε (x) −α(x1 )

Bη (x1 )[D0 (x)]

,

(7.35)

where α(x1 ) is given by (7.3), '

α(x1 )(α(x1 ) + 1) Bη (x1 ) := β(x1 ) + η

(1/(p−1)

,

and, for each x ∈ AR1 ,R2 (x1 + R1 nx1 ) sufficiently close to Γ1 ,



  D0 (x) := dist x, ∂AR1 ,R2 (x1 + R1 nx1 ) = x − (x1 + R1 nx1 ) − R1 .

Now, for each δ ∈ (0, δ0 ], consider the function Lδε defined by Lδε (x) := Lε (x − δnx1 ), By construction,



x ∈ AR1 ,R2 x1 + (R + δ)nx1 .

lim Lδε = Lε . δ↓0

Moreover, for each δ ∈ (0, δ0 ], the restriction Lδε |Ω\Ω0,1

(7.36)

290

J. López-Gómez

is a classical (bounded) subsolution of the singular problem


−
u = λu − aup u=∞

0,1 , in Ω \ Ω on Γ1

since a˜
a. Thus, it follows from Lemma 3.6 that, for each δ ∈ (0, δ0 ], 0,1 . Lδε  L in Ω \ Ω Consequently, passing to the limit as δ ↓ 0, (7.36) implies 0,1 . Lε  L in Ω \ Ω In particular, for each t ∈ (0, R), Lε (x1 − tnx1 )  L(x1 − tnx1 ) and hence, Lε (x1 − tnx1 ) L(x1 − tnx1 )  . −α(x ) 1 Bη (x1 )[dist(x1 − tnx1 , Γ1 )] Bη (x1 )[dist(x1 − tnx1 , Γ1 )]−α(x1 ) On the other hand, by construction, we have that dist(x1 − tnx1 , Γ1 ) = t = D0 (x1 − tnx1 ) and so, passing to the limit as t ↓ 0 in the previous inequality, (7.35) implies 1 − ε  lim inf t↓0

L(x1 − tnx1 ) . Bη (x1 )[dist(x1 − tnx1 , Γ1 )]−α(x1 )

As this inequality holds true for each sufficiently small η > 0 and ε > 0, (7.33) holds. Therefore, by (7.32), the proof of (7.2) is concluded. Condition (7.4) is a direct consequence from (7.2). This concludes the proof of Proposition 7.1. At this stage, the validity of Proposition 7.2 should be apparent.

7.4. Proof of Theorems 2.6 and 2.7 Suppose (2.9), (2.10), and pick λ < σ2 . Subsequently, we set L1 := Lmin  [λ,Ω\Ω

0,1 ]

,

L2 := Lmax  [λ,Ω\Ω

0,1 ]

,

0,1 . D := Ω \ Ω

Theorem 2.6 establishes that L1 = L2 . By construction, we already know that L1  L2 .

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

291

Moreover, thanks to Lemma 5.1, (5.24) and (5.26) imply L1 (x) > 0,

x ∈ D.

L EMMA 7.5. Suppose L1 < L2 . Then, for each x ∈ D, L1 (x) < L2 (x).

(7.37)

P ROOF. We proceed by contradiction. In the contrary case, there exist x0 ∈ D and R > 0 R (x0 ) ⊂ D, L1 (x) < L2 (x) for each x ∈ BR (x0 ) and L1 (y) = L2 (y) for some such that B y ∈ ∂BR (x0 ). Now, consider the problem
−
u = λu − aup in BR (x0 ), (7.38) u = L2 on ∂BR (x0 ). R (x0 ), which contradicts If L1 = L2 on ∂BR (x0 ) then, thanks to Theorem 3.5, L1 = L2 in B L1 < L2 in BR (x0 ). Thus, L1 < L2 on ∂BR (x0 ), and hence, L1 is a positive strict subsolution of (7.38), whose unique solution is L2 . So, thanks to Lemma 3.6, L2 − L1 ≫ 0. Consequently, since L2 (y) − L1 (y) = 0, we have that ∂(L2 − L1 ) (y) < 0, ∂ny

ny :=

y − x0 , R

and therefore, for each sufficiently small t > 0, L2 (y + tny ) < L1 (y + tny ), which contradicts L1  L2 , so concluding the proof of (7.37).



Subsequently, we consider the function q defined through q(x) :=

L1 (x) , L2 (x)

x ∈ D.

(7.39)

As an immediate consequence from the previous properties, the function q is well defined in D and, actually, it is a function of class C 2+µ (D) such that 0 < q(x)  1 for each x ∈ D.

(7.40)

Moreover, thanks to Proposition 7.1, lim q(x1 − tnx1 ) = 1 for each x1 ∈ Γ1 = ∂D, t↓0

(7.41)

and, due to Lemma 7.5, 0 < q(x) < 1 for each x ∈ D

(7.42)

292

J. López-Gómez

if L1 < L2 . The following lemma is the basic technical tool to prove Theorems 2.6 and 2.7.  Then, L1 = L2 . L EMMA 7.6. Suppose q admits a continuous extension Q ∈ C(D).  be a continuous extension of q. Then, thanks to (7.40) and (7.41), P ROOF. Let Q ∈ C(D) 0 < Q = q  1 in D and Q = 1 on ∂D. Moreover, differentiating the identity L1 = QL2 gives −
L1 = −L2
Q − 2∇L2 , ∇Q − Q
L2 , and hence,

p p λL1 − aL1 = −L2
Q − 2∇L2 , ∇Q + Q λL2 − aL2 .

Consequently, dividing by L2 and rearranging terms, we find that −
Q − 2

1

2

p−1 ∇L2 p−1  , ∇Q = aQ L2 − L1
0. L2

Now, for each sufficiently large n ∈ N, say n
n0 , we consider the approximating open sets


1 . Dn := x ∈ D: dist(x, ∂D) > n Thanks to the maximum principle, for each n
n0 , minD n Q must be taken at some point x1n ∈ ∂Dn . Actually, this is a consequence from Theorem 3.1 applied to Q − minD n Q. By construction, the sequence x1n approximates Γ1 = ∂D as n ↑ ∞. Thus, there exists x1 ∈ Γ1 and a subsequence of x1n , again labeled by n, such that lim x1n = x1 .

n↑∞

Therefore,

 1 = Q(x1 ) = lim Q x1n = lim min Q = min Q, n↑∞ D n

n↑∞

 D

and consequently, Q = 1 in D, which concludes the proof. As an immediate consequence from the proof of Lemma 7.6, it is apparent that lim inf q = inf q < lim sup q = sup q = 1 x→x1

 D

x→x1

 D

if L1 < L2 . Moreover, the following uniqueness result holds.



Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

293

L EMMA 7.7. Under the assumptions of Lemma 7.3 and Lemma 7.4, each of the problems (7.9) and (7.21) possesses a unique solution. Moreover, the solution is radially symmetric. P ROOF. Throughout this proof, we consider  Ω ∈ BR (x0 ), AR1 ,R2 (x0 )

and, for sufficiently large n,


1 . Ωn := x ∈ Ω: dist(x, ∂Ω) > n Going back to the proof of Proposition 4.7, it is apparent, by construction, that L1 := Lmin [λ,Ω] and Lmin [λ,Ωn ] are radially symmetric. Thus, min L2 := Lmax [λ,Ω] = lim L[λ,Ωn ] n↑∞

also is radially symmetric. Therefore, the quotient of these solutions, q = L1 /L2 , is radially symmetric as well. On the other hand, adapting the proof of Propositions 7.1 and 7.2 to the domain Ω, it is rather clear that, for each x1 ∈ ∂Ω, lim t↓0

Lj (x1 − tnx1 ) = 1, Bt −α

j ∈ {1, 2},

(7.43)

where α and B are given by (7.11). Hence, lim q(x1 − tnx1 ) = 1 for each x1 ∈ ∂Ω. t↓0

(7.44)

 As q is radially symmetric, due to (7.44), q admits a continuous extension Q ∈ C(Ω), and therefore, adapting Lemma 7.6 to cover the present situation, we obtain that L1 = L2 , which concludes the proof.  Subsequently, in order to refine it, we go back to the proof of Proposition 7.1. By Lemma 7.7, we already know that, for every x1 ∈ Γ1 , each of the auxiliary problems, (7.29) and


−
u = λu − au ˜ p u=∞

in AR1 ,R2 (x1 + R1 nx1 ),

on ∂AR1 ,R2 (x1 + Rnx1 ),

(7.45)

possesses a unique positive solution. Let denote them by Lx01 and Lx01 , respectively. Note that in the proof of Proposition 7.1 the constants δ0 , R, R1 and R2 can be chosen to be independent of x1 ∈ Γ1 since Γ1 is of class C 2 . Subsequently, instead of arguing with the large solutions Lε and Lε whose existence was guaranteed by Lemmas 7.3 and 7.4, we

294

J. López-Gómez

repeat the argument of the proof of Proposition 7.1 using Lx01 and Lx01 . Then, for every η > 0, x1 ∈ Γ1 and t ∈ (0, R), we find that Lx01 (x1 − tnx1 ) L(x1 − tnx1 ) Lx01 (x1 − tnx1 )   B(x1 )t −α(x1 ) B(x1 )t −α(x1 ) B(x1 )t −α(x1 ) and, consequently, x x Bη (x1 ) L01 (x1 − tnx1 ) L(x1 − tnx1 ) Bη (x1 ) L01 (x1 − tnx1 )   . B(x1 ) Bη (x1 )t −α(x1 ) B(x1 ) Bη (x1 )t −α(x1 ) B(x1 )t −α(x1 )

(7.46)

Note that in the special case when α(x1 ) and B(x1 ) are constant the large solutions Lx01 and Lx01 are translations of a fixed profile along each of the points at a distance R from Γ1 , and therefore, by (7.43), lim t↓0

Lx01 (x1 − tnx1 ) Lx01 (x1 − tnx1 ) = 1 = lim t↓0 Bη (x1 )t −α(x1 ) Bη (x1 )t −α(x1 )

(7.47)

uniformly in x1 ∈ Γ1 . Thus, passing to the limit as t ↓ 0 in (7.46) we obtain that Bη (x1 ) L(x1 − tnx1 ) Bη (x1 ) L(x1 − tnx1 )  lim inf  lim sup  t↓0 B(x1 )t −α(x1 ) B(x1 ) B(x1 ) B(x1 )t −α(x1 ) t↓0

(7.48)

uniformly in x1 ∈ Γ1 . Consequently, since lim η↓0

Bη (x1 ) Bη (x1 ) = 1 = lim η↓0 B(x1 ) B(x1 )

uniformly in x1 ∈ Γ1 ,

it follows from (7.48) that lim t↓0

L(x1 − tnx1 ) = 1 uniformly in x1 ∈ Γ1 , B(x1 )t −α(x1 )

(7.49)

which, in particular, entails the continuity of the quotient function q defined in (7.39) and, thanks to Lemma 7.6, concludes the proof of the uniqueness in this special case. In our general setting, the proof is completed if we are able to show that (7.47) is as well satisfied uniformly in x1 ∈ Γ1 . This is easily realized by having a careful look at the proof of Lemma 7.3, which is based upon the construction of the supersolution ψ¯ ε (r) := A + B+

 2 r (R − r)−α , R

B+ := (1 + ε)B,

and the subsolution  2


r −α , ψ ε (r) := max 0, C + B− (R − r) R

B− := (1 − ε)B.

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

295

Rereading the technical details of their constructions, it is rather obvious that the constants A and C can be chosen to be the same for α and β(R), and, hence, B, varying in any compact subset of (0, ∞). Therefore, B+ ψ¯ ε (r) = r↑R B(R − r)−α B lim

and

lim

r↑R

ψ ε (r) B(R − r)−α

=

B− B

for α and β(R) varying in any compact subset of (0, ∞). As α(x1 ) and β(x1 ) are continuous functions of x1 ∈ Γ1 and Γ1 is compact, (7.47) occurs uniformly in x1 ∈ Γ1 , which concludes the proof of Theorem 2.6. The proof of Theorem 2.7 follows by repeating the argument along the remaining component of ∂Ω− . 7.5. Proof of Theorem 2.8 Throughout this subsection we suppose the assumptions of Theorem 2.7 are satisfied. Under these assumptions, Theorem 2.6 can be applied and hence, for each λ ∈ R, (1.9) has a unique large solution in Ω− , denoted by L[λ,Ω− ] , and, for each λ < σ2 , it has a unique 0,1 , denoted by L[λ,Ω\Ω ] . Theorem 2.8 will follow after a selarge solution in Ω \ Ω 0,1 ries of lemmas. The next result shows the strong monotonicity of these large solutions as functions of the parameter λ. L EMMA 7.8. For each λ, µ ∈ R, λ < µ, one has that (a) L[λ,Ω− ] (x) < L[µ,Ω− ] (x) for every x ∈ Ω− ; 0,1 provided µ < σ2 . (b) L[λ,Ω\Ω0,1 ] (x) < L[µ,Ω\Ω0,1 ] (x) for every x ∈ Ω \ Ω P ROOF. By construction, we already know that L[λ,Ω− ] = lim θ[λ,Ω− ,M]

and L[µ,Ω− ] = lim θ[µ,Ω− ,M] .

M↑∞

M↑∞

Moreover, for each M > 0, θ[λ,Ω− ,M] ≪ θ[µ,Ω− ,M]

in Ω− .

Therefore, passing to the limit as M ↑ ∞ gives L[λ,Ω− ]  L[µ,Ω− ]

in Ω− .

Note that, necessarily, L[λ,Ω− ] < L[µ,Ω− ]

in Ω−

(7.50)

since L[λ,Ω− ] = L[µ,Ω− ] implies λ = µ, which contradicts λ < µ. Now, we proceed by contradiction. Suppose part (a) is not satisfied. Then, there exist x0 ∈ Ω− and R > 0 such R (x0 ) ⊂ Ω− , that B L[λ,Ω− ] (x) < L[µ,Ω− ] (x)

for each x ∈ BR (x0 )

(7.51)

296

J. López-Gómez

and L[λ,Ω− ] (y) = L[µ,Ω− ] (y) for some y ∈ ∂BR (x0 ).

(7.52)

Now, consider the auxiliary problem


−
u = λu − aup u = L[µ,Ω− ]

in BR (x0 ), on ∂BR (x0 ).

(7.53)

If L[λ,Ω− ] = L[µ,Ω− ] on ∂BR (x0 ), then, thanks to Theorem 3.5, L[λ,Ω− ] = L[µ,Ω− ] R (x0 ), which contradicts (7.51). Thus, in B L[λ,Ω− ] < L[µ,Ω− ]

on ∂BR (x0 ).

Hence, L[λ,Ω− ] is a positive strict subsolution of (7.53), whose unique solution is L[µ,Ω− ] . So, thanks to Lemma 3.6, L[µ,Ω− ] − L[λ,Ω− ] ≫ 0

R (x0 ). in B

Consequently, it follows from (7.52) that ∂(L[µ,Ω− ] − L[λ,Ω− ] ) (y) < 0, ∂ny

ny :=

y − x0 , R

and therefore, for each sufficiently small t > 0, L[µ,Ω− ] (y + tny ) < L[λ,Ω− ] (y + tny ), which contradicts (7.50). Consequently, part (a) holds true. The previous proof adapts mutatis mutantis to prove part (b).  The next result provides us with the continuity of the large solutions as functions of the parameter λ. L EMMA 7.9. For each x ∈ Ω− the map λ ∈ R → L[λ,Ω− ] (x)

(7.54)

0,1 , the map is continuous. Similarly, for each x ∈ Ω \ Ω λ ∈ (−∞, σ2 ) → L[λ,Ω\Ω0,1 ] (x) is continuous.

(7.55)

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

297

P ROOF. Thanks to Lemma 7.8(a), for each λ ∈ R, the point-wise limit Lλ := lim L[λ−ε,Ω− ] ε↓0

is well defined and finite, since 0 < ε2 < ε1 implies L[λ−ε1 ,Ω− ] (x) < L[λ−ε2 ,Ω− ] (x) < L[λ,Ω− ] (x),

x ∈ Ω− .

Moreover, by the compactness of (−
)−1 , it provides us with a large positive solution of (1.9) in Ω− ; necessarily L[λ,Ω− ] , by uniqueness. Thus, L[λ,Ω− ] = lim L[λ−ε,Ω− ] . ε↓0

Similarly, 0 < ε2 < ε1 implies L[λ+ε1 ,Ω− ] (x) > L[λ+ε2 ,Ω− ] (x) > L[λ,Ω− ] (x),

x ∈ Ω− ,

and consequently, lim L[λ+ε,Ω− ] = L[λ,Ω− ] , ε↓0

which shows the continuity of (7.54). Adapting this argument, the continuity of (7.55) follows readily.  Finally, the relations (2.16) follow from (2.5) and (2.7), by the uniqueness of the underlying large positive solutions, which is guaranteed by Theorems 2.6 and 2.7. This concludes the proof of Theorem 2.8.

8. Relevant bibliography and further results The most pioneering result concerning the positive solutions of (1.8) is by Brezis and Oswald [8], Theorem 1, where it was shown, using variational methods, that (1.8) possesses a positive solution, necessarily unique, if and only if & % & % (8.1) σ −
− a0 (x), Ω < 0 < σ −
− a∞ (x), Ω , where a0 (x) and a∞ (x) are the functions defined through

 a0 (x) := lim λ − a(x)f (x, u) = λ u↓0

and


 −∞ a∞ (x) := lim λ − a(x)f (x, u) = u↑∞ λ

if x ∈ Ω− , if x ∈ Ω \ Ω− .

298

J. López-Gómez

Although it is rather obvious that the first inequality of (8.1) becomes λ > σ0 , it is far from easy to realize why the second inequality should become λ < σ1 = σ [−
, Ω0,1 ], in order to provide us with the existence interval λ ∈ (σ0 , σ1 ) given by Theorem 2.3 here. Actually, this is an easy consequence from an extremely sharp result going back to [41], according with it one has that % & lim σ −
+ af (·, u), Ω = σ [−
, Ω0,1 ] = σ1 .

u↑∞

(8.2)

Indeed, due to (8.2), (8.1) becomes into σ0 < λ < σ1 .

(8.3)

Another significant pioneering contribution was done by Ouyang [62], as a part of his Ph.D. Dissertation on Yamabe’s problem [68] under the supervision of W.-M. Ni. Combining global continuation with the existence of uniform a priori bounds in compact subsets of λ ∈ (−∞, σ1 ), Ouyang [62] established that, under assumption (2.9), condition (8.3) characterizes the existence of positive solutions of (1.8) and that lim θ[λ,Ω] L2 (Ω) = ∞.

λ↑σ1

(8.4)

Later, Ambrosetti and Gámez [4] and del Pino [26] could improve (8.4) up to get lim θ[λ,Ω] L∞ (Ω) = ∞,

λ↑σ1

(8.5)

though in none of these works any explicit mention was made towards the problem of analyzing the different role played by each of the components of a −1 (0), Ω0,1 and Ω0,2 , in ascertaining the dynamics of (1.1). The first general characterizations of the existence of positive solutions for a generalized class of logistic models including the basic prototype (1.8) were obtained by Fraile et al. [29] as a consequence from the method of sub- and supersolutions. In [29], instead of −
, a general elliptic operator of second-order subject to rather general boundary conditions was considered. At the present stage of this monograph, the reader should be fully convinced that the method of sub- and supersolutions is the best available method to deal with generalized sublinear problems, since the supersolution itself provides us with the exact shape profile of the positive solutions of (1.8) for λ > σ0 separated away from σ0 ; for λ ∼ σ0 , the solution looks like the subsolution of the problem. Besides this tremendous advantage, the method of sub- and supersolutions provides us simultaneously with the global attractive character of the maximal nonnegative solution of (1.8) within the range λ ∈ (−∞, σ1 ). Undoubtedly, [29] has been a milestone for the further development of the theory that we have exposed in this monograph. Actually, the problem of the analysis of the asymptotic behavior of the solutions of (1.1) for λ
σ1 was originally addressed in [29],

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

299

where it was shown that the solutions of (1.1) must be unbounded, as t ↑ ∞, within Ω0,1 . Most of the results of [29] were obtained during the academic year 1993–1994, as a result of a question raised by P. Koch and S. Merino to the author during his first visit to Zürich University in June 1993. The paper was submitted for publication on November 1994. According to the proof of Theorem 2.3 in Section 5, for each λ ∈ (σ0 , σ1 ), the positive supersolution of (1.8) has the form κΦ, where Φ is the function defined in (5.7) for a sufficiently large κ > 1 that blows up as λ ↑ σ1 . Quite strikingly, Φ is the supersolution constructed in [41] to prove (8.2). This crucial bisociation, besides sharpening the paradigmatic theorem of Brezis and Oswald [8], enjoys a huge interest in its own right, because of its number of applications in Mathematical Biology (cf. [11,12,42], and the references therein), in the study of linear weighted boundary value problems (cf. [41] and [43]), and in the semiclassical analysis of linear oscillators at degenerate wells (cf. [22], where it was used to solve some classical open problems proposed by Simon [66]). Bisociation. . . a word coined by the political and scientific writer A. Koestler for designating unexpected – sharply hidden – connections between a priori unrelated separated fields (cf. [9]). According to Theorem 2.3, the mapping λ → θ[λ,Ω] is differentiable and strictly pointwise increasing. Moreover, by (2.6), θ[λ,Ω] bifurcates from u = 0 at λ = σ0 and, due to (2.7), 0,1 \ ∂Ω. lim θ[λ,Ω] (x) = ∞ for each x ∈ Ω

λ↑σ1

(8.6)

The fact that (8.6) occurs in Ω0,1 goes back to [45], Theorem 2.4, which was submitted for publication on August 1996, thought it appeared in 2000. The proof given in Section 5 is the original one given in [45]. It should be noted that (8.6) extremely sharpens (8.5). In Figure 10 we have represented θ[λ,Ω] (x) versus λ ∈ (σ0 , σ1 ) for a given x ∈ Ω0,1 . Thanks to (8.6), the diagram exhibits a bifurcation from infinity at λ = σ1 . In Figure 10 stable solutions are indicated by solid lines, and unstable solutions by dashed lines. The

Fig. 10. The steady-states of (1.8).

300

J. López-Gómez

state u = 0 loses stability as λ crosses σ0 , such stability being gained by θ[λ,Ω] . By Theorem 2.4, local stability actually entails global attractiveness. It is remarkable the significant difference between the diagram shown by Figure 10 and the diagram of the classical positive solutions that has been already inserted in Figure 2, where the curve λ → θ[λ,Ω] (x) is bounded at λ = σ1 . The difference coming from the fact that, thanks to (2.7), for each 0,1 one has that x ∈Ω \Ω lim θ[λ,Ω] (x) = Lmin  [σ ,Ω\Ω 1

λ↑σ1

0,1 ]

(x) < ∞,

(8.7)

in severe contrast with (8.6). Bifurcation diagrams might drastically change according to the magnitude chosen to represent it, as it actually occurs here in, of course. Subsequently, we consider the compact set K defined by


1 K := x ∈ Ω: dist(x, ∂Ω)
n for a sufficiently large n ∈ N (see Figure 1). Thanks to the Harnack inequality, for each λ ∈ (σ0 , σ1 ) there exists a constant Cλ > 0 such that max θ[λ,Ω]  Cλ min θ[λ,Ω] . K

K

(8.8)

Thus, if there are ε > 0 and C > 0 for which sup σ1 −ε 1, in order to obtain the following result, whose proof is omitted here. T HEOREM 8.2. Suppose 1 < m < p. Then (8.15) possesses a positive solution (unique) if and only if λ > 0. Moreover, if we denote it by W[λ,Ω,m] , then, setting m , θ[λ,Ω,m] := W[λ,Ω,m]

the following properties are satisfied: 1. limm↓1 θ[λ,Ω,m] = 0 if 0 < λ  σ0 . 2. limm↓1 θ[λ,Ω,m] = θ[λ,Ω] if σ0 < λ < σ1 . 3. limm↓1 θ[λ,Ω,m] = Mmin  ] if σ1  λ < σ2 . [λ,Ω\Ω 0,1

4.

limm↓1 θ[λ,Ω,m] = Mmin [λ,Ω− ]

if λ
σ2 .

Metasolutions have also shown to be extremely relevant in analyzing the dynamics of very large classes of reaction diffusion systems (cf. [27,49] and [21]), and of very general

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

305

classes of indefinite superlinear parabolic problems, where the weight function a(x) is allowed to change sign within Ω. In order to sketch how metasolutions arise in the context of indefinite superlinear problems, suppose a satisfies the following 0,1 ∪ ∂Ω0,2 , a −1 (0) = Ω

 a −1 (0, ∞) = Ω− ,

 a −1 (−∞, 0) = Ω0,2 , (8.16)

and denote by a + and a − the positive and negative parts of a, a = a+ − a−, whose supports are Ω− and Ω0,2 , respectively. Then, by bringing together the theory developed in this monograph with the techniques and results of López-Gómez and Quittner [55] and Quittner and Simondon [64], one can get the following result, which is a sort of summary of the main findings of [52]. T HEOREM 8.3. Suppose (8.16), (2.9), λ > σ1 , and u0 > 0 satisfies −
u0  λu0 − a + up . Then, the following properties are satisfied: 1. For each λ ∈ (σ0 , σ1 ), there exists ε(λ) > 0 such that a − C (Ω0,2 ) < ε(λ) implies lim u[λ,Ω] (·, t; u0 ) = θ[λ,Ω] ,

t↑∞

where θ[λ,Ω] stands for the minimal positive solution of (1.8), whose existence is guaranteed by Amann and López-Gómez [3]. 2. For each λ ∈ (σ0 , σ1 ), there exists κ(λ) > ε(λ) such that u[λ,Ω] (·, t; u0 ) blows up 0,2 in a finite time if a − C (Ω0,2 ) > κ(λ). Moreover, the blow up is complete in Ω if p − 1 > 0 is sufficiently small. Furthermore, in this case, u[λ,Ω] (·, t; u0 ) approxi mates Mmin  ] as t ↑ ∞, though, in Ω0,2 , infinity is reached in a finite time. [λ,Ω\Ω 0,2

3. For each λ ∈ [σ1 , σ2 ), there exists ε(λ) > 0 such that a − C (Ω0,2 ) < ε(λ) implies lim u[λ,Ω] (·, t; u0 ) = Mmin  [λ,Ω\Ω

t↑∞

0,1 ]

,

where Mmin 0,1 ] stands for the minimal metasolution of (1.9) supported [λ,Ω\Ω 0,1 .  in Ω \ Ω0,1 . Note that a(x) changes sign in Ω \ Ω 4. For each λ ∈ [σ1 , σ2 ), there exists κ(λ) > ε(λ) such that u[λ,Ω] (·, t; u0 ) blows up 0,2 in a finite time if a − C (Ω0,2 ) > κ(λ). Moreover, the blow up is complete in Ω if p − 1 > 0 is sufficiently small. Furthermore, in this case, u[λ,Ω] (·, t; u0 ) approxi mates Mmin [λ,Ω− ] as t ↑ ∞, though, in Ω0,2 , infinity is reached in a finite time.

306

J. López-Gómez

5. For each λ
σ2 , u[λ,Ω] (·, t; u0 ) blows up in a finite time. Moreover, the blow up 0,2 if p − 1 > 0 is sufficiently small. Furthermore, in this case, is complete in Ω  u[λ,Ω] (·, t; u0 ) approximates Mmin [λ,Ω− ] as t ↑ ∞, though, in Ω0,2 , infinity is reached in a finite time. Consequently, though metasolutions arose in analyzing the most paradigmatic model of Population Dynamics, they seem to be crucial to understand the role of spatial heterogeneities in wide areas of Science and Technology. Undoubtedly, within the next few years metasolutions will play a significant role in the modern theory of Partial Differential Equations, where using spatial heterogeneities is imperative in order to get more realistic models.

Acknowledgments This research has been supported by the Ministry of Education and Science of Spain under grant REN2003-00707. The author expresses his deepest gratitude to the editors for giving him the opportunity to tidy up in this monograph part of the materials that he has studied during the past 10 years. Also, he wants to dedicate this paper to all those mathematicians that have supported him – in very diverse ways – during less optimistic times and, very specially, to Professors H. Amann, J.K. Hale and J. Mawhin.

References [1] H. Amann, On the existence of positive solutions of nonlinear elliptic boundary value problems, Indiana Univ. Math. J. 21 (1971), 125–146. [2] H. Amann, Fixed point equations and nonlinear eigenvalue problems in ordered Banach spaces, SIAM Rev. 18 (1976), 620–709. [3] H. Amann and J. López-Gómez, A priori bounds and multiple solutions for superlinear indefinite elliptic problems, J. Differential Equations 146 (1998), 336–374. [4] A. Ambrosetti and J.L. Gámez, Branches of positive solutions for some semilinear Schrödinger equations, Math. Z. 224 (1997), 347–362. [5] C. Bandle and M. Marcus, Sur les solutions maximales de problémes elliptiques nonlinéaires: bornes isopérimetriques et comportement asymptotique, C. R. Acad. Sci. Paris Sér. I 311 (1990), 91–93. [6] C. Bandle and M. Marcus, Large solutions of semilinear elliptic equations: Existence, uniqueness and asymptotic behavior, J. Anal. Math. 58 (1991), 9–24. [7] C. Bandle and M. Marcus, On second order effects in the boundary behavior of large solutions of semilinear elliptic problems, Differential Integral Equations 11 (1998), 23–34. [8] H. Brezis and L. Oswald, Remarks on sublinear elliptic equations, Nonlinear Anal. 10 (1986), 55–64. [9] F. Browder, Reflection on the future of mathematics, Notices Amer. Math. Soc. 49 (2002), 658–662. [10] S. Cano-Casanova and J. López-Gómez, Properties of the Principal eigenvalues of a general class of non-classical mixed boundary value problems, J. Differential Equations 178 (2002), 123–211. [11] S. Cano-Casanova and J. López-Gómez, Permanence under strong aggressions in possible, Ann. Inst. H. Poincaré Anal. Non Linéaire 20 (2003), 999–1041. [12] S. Cano-Casanova, J. López-Gómez and M. Molina-Meyer, Permanence through spatial segregation in heterogeneous competition, Proc. 9th IEEE Internat. Conf. Methods and Models in Automation and Robotics, R. Kaszy´nski, ed., Institute of Control Engineering of the Technical University of Szczecin, Szczecin (2003), 123–130.

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

307

[13] S. Cano-Casanova, J. López-Gómez and M. Molina-Meyer, Isolas: Compact solution components separated away from a given equilibrium curve, Hiroshima Math. J. 34 (2004), 177–199. [14] F.C. Cirstea and V. Radulescu, Existence and uniqueness of blow-up solutions for a class of logistic equations, Commun. Contemp. Math. 4 (2002), 559–586. [15] F.C. Cirstea and V. Radulescu, Uniqueness of the blow-up boundary solution of logistic equation with absorption, C. R. Acad. Sci. Paris Sér. I 335 (2002), 447–452. [16] F.C. Cirstea and V. Radulescu, Asymptotics for the blow-up boundary solution of the logistic equation with absorption, C. R. Acad. Sci. Paris Sér. I 336 (2003), 231–236. [17] F.C. Cirstea and V.D. Radulescu, Solutions with boundary blow-up for a class of nonlinear elliptic problems, Houston J. Math. 29 (2003), 821–829. [18] R. Courant and D. Hilbert, Methods of Mathematical Physics, Vol. I, Wiley, New York (1962). [19] M.G. Crandall and P.H. Rabinowitz, Bifurcation from simple eigenvalues, J. Funct. Anal. 8 (1971), 321–340. [20] M.G. Crandall and P.H. Rabinowitz, Bifurcation, perturbation from simple eigenvalues and linearized stability, Arch. Ration. Mech. Anal. 52 (1973), 161–180. [21] E.N. Dancer and Y. Du, Effect of certain degeneracies in the predator–prey model, SIAM J. Math. Anal. 34 (2002), 292–314. [22] E.N. Dancer and J. López-Gómez, Semiclassical analysis of general second order elliptic operators on bounded domains, Trans. Amer. Math. Soc. 352 (2000), 3723–3742. [23] M. Delgado, J. López-Gómez and A. Suárez, Nonlinear versus linear diffusion. From classical solutions to metasolutions, Adv. Differential Equations 7 (2002), 1101–1124. [24] M. Delgado, J. López-Gómez and A. Suárez, Characterizing the existence of large solutions for a class of sublinear problems with nonlinear diffusion, Adv. Differential Equations 7 (2002), 1235–1256. [25] M. Delgado, J. López-Gómez and A. Suárez, Singular boundary value problems of a porous media logistic equation, Hiroshima Math. J. 34 (2004), 57–80. [26] M.A. del Pino, Positive solutions of a semilinear elliptic equation on a compact manifold, Nonlinear Anal. 22 (1994), 1423–1430. [27] Y. Du, Effects of a degeneracy in the competition model. Classical and generalized steady-state solutions, J. Differential Equations 181 (2002), 92–132. [28] Y. Du and Q. Huang, Blow-up solutions for a class of semilinear elliptic and parabolic equations, SIAM J. Math. Anal. 31 (1999), 1–18. [29] J.M. Fraile, P. Koch-Medina, J. López-Gómez and S. Merino, Elliptic eigenvalue problems and unbounded continua of positive solutions of a semilinear equation, J. Differential Equations 127 (1996), 295–319. [30] J.E. Furter and J. López-Gómez, Diffusion mediated permanence problem for an heterogeneous Lotka– Volterra competition model, Proc. Roy. Soc. Edinburgh Sect. A 127 (1997), 281–336. [31] J. García-Melián, R. Gómez-Reñasco, J. López-Gómez and J.C. Sabina de Lis, Point-wise growth and uniqueness of positive solutions for a class of sublinear elliptic problems where bifurcation from infinity occurs, Arch. Ration. Mech. Anal. 145 (1998), 261–289. [32] J. García-Melián, R. Letelier-Albornoz and J.C. Sabina de Lis, Uniqueness and asymptotic behavior for solutions of semilinear problems with boundary blow-up, Proc. Amer. Math. Soc. 129 (2001), 3593–3602. [33] D. Gilbarg and N.S. Trudinger, Elliptic Partial Differential Equations of Second Order, Springer, Berlin (2001). [34] R. Gómez-Reñasco, The effect of varying coefficients in semilinear elliptic boundary value problems. From classical solutions to metasolutions. Ph.D. Dissertation, Universidad de La Laguna, Tenerife (1999). [35] R. Gómez-Reñasco and J. López-Gómez, The effect of varying coefficients on the dynamics of a class of superlinear indefinite reaction diffusion equations, J. Differential Equations 167 (2000), 36–72. [36] R. Gómez-Reñasco and J. López-Gómez, On the existence and numerical computation of classical and non-classical solutions for a family of elliptic boundary value problems, Nonlinear Anal. 48 (2002), 567–605. [37] J. Hadamard, Mémoires sur le problème d’analyse relatif à l’equilibre des plaques élastiques encastrées, Mém. Acad. Sci. Paris 39 (1908), 128–259. [38] J.B. Keller, On solutions of
u = f (u), Comm. Pure Appl. Math. X (1957), 503–510. [39] V.A. Kondratiev and V.A. Nikishin, Asymptotics, near the boundary, of a solution of a singular boundary value problem for a semilinear elliptic equation, Differential Equations 26 (1990), 345–348.

308

J. López-Gómez

[40] C. Loewner and L. Nirenberg, Partial differential equations invariant under conformal or projective transformations, Contributions to Analysis, Academic Press, New York (1974), 245–272. [41] J. López-Gómez, On linear weighted boundary value problems, Partial Differential Equations: Models in Physics and Biology, Akademie-Verlag, Berlin (1994), 188–203. [42] J. López-Gómez, Permanence under strong competition, Dynamical Systems and Applications, World Scientific Series in Applied Analysis, Vol. 4, World Scientific, Singapore (1995), 473–488. [43] J. López-Gómez, The maximum principle and the existence of principal eigenvalues for some linear weighted boundary value problems, J. Differential Equations 127 (1996), 263–294. [44] J. López-Gómez, Large solutions, metasolutions, and asymptotic behavior of the regular positive solutions of sublinear parabolic problems, Electron. J. Differ. Equ. Conf. 05 (2000), 135–171. [45] J. López-Gómez, Varying bifurcation diagrams of positive solutions for a class of indefinite superlinear boundary value problems, Trans. Amer. Math. Soc. 352 (2000), 1825–1858. [46] J. López-Gómez, Spectral Theory and Nonlinear Functional Analysis, Chapman & Hall/CRC Res. Notes Math., Vol. 426, Chapman & Hall/CRC, Boca Raton, FL (2001). [47] J. López-Gómez, Approaching metasolutions by classical solutions, Differential Integral Equations 14 (2001), 739–750. [48] J. López-Gómez, Classifying smooth supersolutions in a general class of elliptic boundary value problems, Adv. Differential Equations 8 (2003), 1025–1042. [49] J. López-Gómez, Coexistence and metacoexistence states in competing species models, Houston J. Math. 29 (2003), 485–538. [50] J. López-Gómez, The boundary blow-up rate of large solutions, J. Differential Equations 195 (2003), 25–45. [51] J. López-Gómez, Dynamics of parabolic equations. From classical solutions to metasolutions, Differential Integral Equations 16 (2003), 813–828. [52] J. López-Gómez, Global existence versus blow-up in superlinear indefinite parabolic problems, Sciantiae Mathematicae Japonicae 17 (2005), 449–472. [53] J. López-Gómez and M. Molina-Meyer, The maximum principle for cooperative weakly coupled elliptic systems and some applications, Differential Integral Equations 7 (1994), 383–398. [54] J. López-Gómez and M. Molina-Meyer, Bounded components of positive solutions of abstract fixed point equations: mushrooms, loops and isolas, J. Differential Equations 209 (2005), 416–441. [55] J. López-Gómez and P. Quittner, Complete and energy blow-up in indefinite superlinear parabolic problems, Dicrete Contin. Dyn. Syst., to appear. [56] J. López-Gómez and J.C. Sabina de Lis, First variations of principal eigenvalues with respect to the domain and point-wise growth of positive solutions for problems where bifurcation from infinity occurs, J. Differential Equations 148 (1998), 47–64. [57] M. Marcus and L. Véron, Uniqueness and asymptotic behavior of solutions with boundary blow-up for a class of nonlinear elliptic equations, Ann. Inst. H. Poincaré Anal. Non Linéaire 14 (1997), 237–274. [58] M. Marcus and L. Véron, The boundary trace and generalized boundary value problem for semilinear elliptic equations with coercive adsorption, Comm. Pure Appl. Math. 56 (2003), 689–731. [59] J. Mawhin, Les héritiers de Pierre-François Verhulst: une population dynamique, Acad. Roy. Belg. Bull. Cl. Sci. 13 (2003), 349–378. [60] J. Mawhin, The legacy of P.F. Verhulst and V. Volterra in Population Dynamics, The First 60 Years of Nonlinear Analysis of Jean Mawhin, M. Delgado, J. López-Gómez, R. Ortega, A. Suárez, eds, World Scientific, Singapore (2004), 147–160. [61] R. Osserman, On the inequality
u
f (u), Pacific J. Math. 7 (1957), 1641–1647. [62] T. Ouyang, On the positive solutions of semilinear equations
u + λu − hup = 0 on the compact manifolds, Trans. Amer. Math. Soc. 331 (1992), 503–527. [63] M.H. Protter and H.F. Weinberger, Maximum Principles in Differential Equations, Prentice-Hall, Englewood Cliffs, NJ (1967). [64] P. Quittner and F. Simondon, A priori bounds and complete blow-up of positive solutions of indefinite superlinear parabolic problems, Preprint (2004). [65] D. Sattinger, Topics in Stability and Bifurcation Theory, Lectures Notes in Math., Vol. 309, SpringerVerlag, Berlin/New York (1973).

Metasolutions: Malthus versus Verhulst in Population Dynamics. A dream of Volterra

309

[66] B. Simon, Semiclassical analysis of low lying eigenvalues, I. Non-degenerate minima: Asymptotic expansions, Ann. Inst. H. Poincaré Sér. A XXXVIII (1983), 12–37. [67] L. Véron, Semilinear elliptic equations with uniform blow up on the boundary, J. Anal. Math. 59 (1992), 231–250. [68] H. Yamabe, On a deformation of Riemaniann structures on compact manifolds, Osaka Math. J. 12 (1960), 21–37.

This page intentionally left blank

CHAPTER 5

Elliptic Problems with Nonlinear Boundary Conditions and the Sobolev Trace Theorem Julio D. Rossi Departamento de Matemática, Universidad de Buenos Aires, 1428 Buenos Aires, Argentina E-mail: [email protected]

Contents 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2. Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3. Existence results for an elliptic problem with nonlinear boundary conditions. A variational approach 4. Problems in RN + . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5. Yamabe problem on manifolds with boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6. Dependence of the best Sobolev trace constant on the domain . . . . . . . . . . . . . . . . . . . . . . 7. Symmetry of extremals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8. Behavior of the best Sobolev trace constant and extremals in domains with holes . . . . . . . . . . . 9. The Sobolev trace embedding with the critical exponent . . . . . . . . . . . . . . . . . . . . . . . . . 10. Dependence of the best Sobolev trace constant on the exponents . . . . . . . . . . . . . . . . . . . . 11. Elliptic systems with nonlinear boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . 12. Other results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . .

313 316 322 342 344 346 360 365 367 375 378 395 401 401

Abstract This work constitutes a short survey of the subject of elliptic partial differential equations with nonlinear boundary conditions. We will focus especially on the relevance of the Sobolev trace theorem in the analysis of this kind of problems. We will also describe some of the techniques employed when dealing with such a kind of problems.

Keywords: Elliptic problems, Nonlinear boundary conditions, Sobolev inequalities MSC: 35J65, 35J50, 35J55

HANDBOOK OF DIFFERENTIAL EQUATIONS Stationary Partial Differential Equations, volume 2 Edited by M. Chipot and P. Quittner © 2005 Elsevier B.V. All rights reserved 311

Nonlinear boundary conditions

313

1. Introduction In this chapter we provide a survey of mathematical results concerning solutions of elliptic problems with nonlinear boundary conditions. We mean solutions of problems that can be written in the general form


Lu = f (u) Bu = g(u)

in Ω, on ∂Ω.

(1.1)

Here, and through this work, Ω is a bounded domain in RN with smooth boundary ∂Ω, ∂ Lu is a second-order elliptic operator and Bu is a boundary condition that involves ∂ν , the outer normal derivative. Of importance in the study of boundary value problems for differential operators in Ω are the Sobolev spaces and inequalities. Hence, Sobolev inequalities and their optimal constants is a subject of interest in the analysis of PDEs and related topics. It has been widely studied in the past by many authors and is still an area of intensive research. See, for instance, the book [15] and the survey [50] for recent developments in this field. When analyzing problems with nonlinear boundary conditions like (1.1) it turns out that among the Sobolev embeddings the Sobolev trace theorem plays a fundamental role. Also one is lead to the study of nonlinear boundary conditions when one tries to find out properties of the Sobolev trace best constant. Our main interest here is to look closely at this relation between nonlinear boundary conditions and the Sobolev trace theorem. Another motivation to study problems with nonlinear boundary conditions comes from geometry. One is lead to nonlinear boundary conditions when one performs a description of conformal deformations on Riemannian manifolds with boundary. Look at the results of Cherrier and Escobar [29,54,55]. In [53] and [105] a geometric problem in the halfspace RN + with nonlinear boundary conditions is studied. Also, nonlinear boundary conditions appear in a rather natural way in some physical models. For example, problem (1.1) can be thought of as a model for heat propagation. In this case u stands for the temperature and the normal derivative ∂u ∂ν that appears in the boundary condition B(u) represents the heat flux. Hence the boundary condition represents a nonlinear radiation law at the boundary. This kind of boundary conditions appear also in combustion problems when the reaction happens only at the boundary of the container, for example, because of the presence of a solid catalyzer, see [99] for a justification. Eigenvalue problems with the eigenvalue placed at the boundary condition, ∂u ∂ν = λu, are studied since the pioneering work of Steklov, see [115]. There are works that consider fully nonlinear equations like F (x, u, ∇u, D 2 u) = 0 in Ω with nonlinear boundary conditions, H (x, u, ∇u) = 0 on ∂Ω, where H is assumed to be strictly increasing with respect to ∇u in the normal direction to ∂Ω at x. See, for example, [16] and [89] where viscosity solutions are considered. There are also papers that deal with higher-order equations, for example in [72] a fourth-order problem is considered. However, to simplify the exposition, we will be only concerned with second-order problems like (1.1).

314

J.D. Rossi

Elliptic problems, Lu = f (u), with Dirichlet boundary conditions, u|∂Ω = 0, have been widely treated in the literature, see the survey [44] and references therein. Many of the known results that hold for Dirichlet boundary conditions have an analogous counterpart when dealing with nonlinear boundary conditions. However many times the proofs are different. We will try to emphasize the differences and similitudes between both types of boundary conditions. Therefore, a strong motivation to study elliptic problems with nonlinear boundary conditions is to see how the well-established theory for the Dirichlet problem extends to other situations (as nonlinear boundary conditions) and to develop new ideas and methods when the available theory is not applicable. These developments may have consequences for other related problems that do not involve necessarily nonlinear boundary conditions. We have to mention that there is a large amount of literature dealing with parabolic problems with nonlinear boundary conditions. In the last years there has been an increasing interest in the study of blow-up due to reaction at the boundary, both for scalar problems and for systems, see, for example, the surveys [34,77] and references therein. Often parabolic results are related to elliptic results. For example, when every positive solution of a parabolic problem blows up there cannot exist any positive stationary solution. Hence we have a nonexistence result of positive solutions for the elliptic problem in this case, see [30], [31] and [81] for such type of results. The stability properties of a stationary solution is also a problem to deal with, see, for example, [31] and [39]. Moreover, many times regularity results for elliptic and parabolic problems are related, see [4]. The References do not escape the usual rule of being incomplete. In general, we have listed those papers which are more close to the topics discussed here. But, even for those papers, the list is far from being exhaustive and we apologize for omissions. Organization of the chapter. The rest of the chapter contains eleven sections. They are organized by subject, however many times there are relations between them that we will try to outline. In some cases we will provide full proofs (or at least sketch the main arguments) in order to give the reader an idea of the involved techniques. In each section a change of subject or problem will be marked with . Section 2. In this section we state some preliminaries relating the best Sobolev trace constant with problems with nonlinear boundary conditions and give some ideas about regularity results from J. García-Azorero, I. Peral and the author that can be found in [81]. Section 3. In this section we see how to adapt usual variational techniques to deal with nonlinear boundary conditions. We follow ideas from M. Chipot, M. Fila, P. Quittner, J. García-Azorero, I. Peral, K. Umezu, J. Fernández Bonder and the author, see [31], [32], [71], [81] and [121]. Section 4. We collect some results in the half-space, RN + , proved by B. Hu, X. Cabre, M. Chipot, M. Chlebik, I. Shafrir, M. Fila, Y. Park, W. Reichel, J. Sola-Morales, S. Terracini, Y. Li and M. Zhu, see [24], [30], [33], [87], [93], [106] and [117]. Section 5. Here we discuss about the results of A. Ambrosetti, Y. Li, A. Malchiodi, V. Felli, M. Ahmedou and J.F. Escobar, which have a strong geometrical motivation, the socalled Yamabe problem for manifolds with boundary. In this geometrical problem the critical exponent for the Sobolev trace embedding appears. See [5], [53], [54], [55] and [60].

Nonlinear boundary conditions

315

Section 6. In this section we deal with the dependence of the best Sobolev trace constant on the domain for a subcritical exponent. We focus on families of domains obtained by expanding or contracting a fixed domain and obtain the asymptotic behavior of the best Sobolev trace constant. Also we prove that the first eigenvalue of an associated nonlinear Steklov-like eigenvalue problem is isolated and simple. We collect results of C. Flores, M. del Pino, S. Martinez, J. Fernández Bonder and the author, [46], [66], [74] and [100]. Section 7. We look at symmetry of the extremals for the Sobolev trace embedding. In particular, we study when the extremals are radial if Ω is the ball of radius R, B(0, R). We report on the results of O. Torne, E. Lami-Dozo, J. Fernández Bonder and the author, [65] and [92]. Section 8. In this section we consider the trace of functions that belong to a Sobolev space and vanish over some subset of Ω. We look at the problem of optimizing the best Sobolev trace constant when varying the subset where the involved functions vanish keeping its area fixed. We follow ideas of J. Fernández Bonder, N. Wolanski and the author, [76]. Section 9. Here we look at the Sobolev trace embedding with the critical exponent. The results presented in this section are mainly due to E. Abreu, P. Carriao, O. Miyagaki, D. Pierotti, S. Terracini, F. Demengel, M. Motron, M. Chlebik, M. Fila, W. Reichel, F. Andreu, J. Mazon, J. Fernández Bonder and the author, [1], [11], [35], [47], [75], [104], [108] and [109]. See also Section 5 for other results that involve the critical exponent. Section 10. Now we study the dependence of the Sobolev trace constant on the exponents involved. We rely on results of R. Ferreira, J. Fernández Bonder and the author, [64]. Section 11. In this section we collect results concerning elliptic systems with nonlinear boundary conditions, from M. Schechter, W. Zou, S. Li, J. Fernández Bonder, S. Martinez and the author, see [67], [68], [69], [70], [112] and [125]. Section 12. Finally we collect other results for problems with nonlinear boundary conditions, concerning maximum/antimaximum principle, isoperimetric inequalities, selfsimilar profiles for parabolic problems with blow-up, free boundaries, equations involving maximal monotone graphs and their relation with semigroup theory, resonance problems, the Fuˇcik spectrum at the boundary, etc. In this section we do not provide any proofs and refer for details to the papers of F. Andreu, D. Arcoya, Ph. Benilan, F. Brock, M. Crandall, J. Davila, S. Martinez, J.M. Mazon, M. Montenegro, P. Sacks, S. Segura de Leon, J. Toledo and the author. See [12], [14], [19], [23], [41], [42], [101] and [102]. Notations. We end the Introduction fixing some of the notation that will be used in the following sections. Along this chapter there are two measures involved, the usual Lebesgue measure in Ω ⊂ RN and the surface measure on ∂Ω. With dx and dσ we denote the corresponding N - and (N −1)-dimensional measures. Also we will use the notation |A| for the measure of the set A in its corresponding dimension, that is, if A is a set of dimension r, |A| stands for the r-dimensional measure of A. We will call the characteristic function of the set A as χA . With p∗ = p(N − 1)/(N − p) and p ∗ = pN/(N − p) we denote the critical exponents for the Sobolev trace embedding W 1,p (Ω) → Lq (∂Ω) and the Sobolev embedding W 1,p (Ω) → Lr (Ω). S(Ω, p, q) stands for the best Sobolev trace constant, see Section 2. Remark that we make explicit the dependence of the constant on the domain and on the involved exponents. This dependence will be analyzed throughout this work.

316

J.D. Rossi

2. Preliminaries Let us look for a simple example where problems with nonlinear boundary conditions appear in a very natural way. To this end, let us first remark that in the study of elliptic and parabolic partial differential equations the Sobolev spaces are a very useful and versatile tool for their analysis. For general references on Sobolev spaces we cite [2]. Recall that the Sobolev space H 1 (Ω) is defined as the space of L2 (Ω) functions with weak first derivatives that also belong to L2 (Ω). In H 1 (Ω) we have a Hilbert space structure. We consider the usual norm that comes from the inner product,   u, v = ∇u∇v dx + uv dx, Ω

Ω

that is, uH 1 (Ω) =



Ω

2

|∇u| dx +



Ω

2

|u| dx

1/2

.

 ⊂ H 1 (Ω)) we can define the restriction to the Given a smooth function (e.g., u ∈ C 1 (Ω) boundary u|∂Ω . It turns out that this restriction operator can be extended from smooth functions to H 1 (Ω) giving a linear continuous operator from H 1 (Ω) to Lr (∂Ω), if 1  r  2∗ = 2(N − 1)/(N − 2), T : H 1 (Ω) → Lr (∂Ω). This result is the very well-known Sobolev trace theorem. See, for example, [2]. The norm of this operator is given by  S(Ω, 2, r) = inf u2H 1 (Ω) ; u ∈ H 1 (Ω) with u2Lr (∂Ω) = 1  2 2 Ω|∇u| + |u| dx = inf . (2.1) r 2/r u∈H 1 (Ω)\H01 (Ω) ( ∂Ω |u| dσ )

This value S(Ω, 2, r) is known as the best Sobolev trace constant. For values of r < 2∗ (subcritical values) we have that the trace operator is a compact operator, therefore an easy compactness argument proves that there exist extremals, that is, functions in H 1 (Ω) where the norm is attained. These extremals turn out to be weak solutions of

u = u in Ω, (2.2) ∂u r−2 u on ∂Ω, ∂ν = λ|u| where λ is a Lagrange multiplier. For a weak solution of (2.2) we understand a function u ∈ H 1 (Ω) that verifies    ∇u∇v dx + uv dx = λ|u|r−2 uv dσ Ω

Ω

∂Ω

(2.3)

Nonlinear boundary conditions

317

for every test function v ∈ H 1 (Ω). Remark that in (2.3) two measures (dx, the volume measure and dσ , the surface measure) are involved. If we assume the normalization of the extremal given by uLr (∂Ω) = 1, taking v = u in (2.3), we get that the Lagrange multiplier is related to the best Sobolev trace constant given by (2.1). It holds, λ = S(Ω, 2, r). Problem (2.2) has associated an energy functional   |u|r 1 |∇u|2 + |u|2 dx − λ dσ. (2.4) F(u) = 2 Ω r ∂Ω In fact, critical points of (2.4) in H 1 (Ω) are weak solutions of the problem (2.2) in the weak sense (2.3). An important case is when r = 2. In this case (2.2) becomes the linear eigenvalue problem

u = u in Ω, ∂u ∂ν = λu on ∂Ω. Using the compactness of the embedding H 1 (Ω) ֒→ L2 (∂Ω) it is obtained that there exists a sequence of eigenvalues λn → ∞ and the first one corresponds to the best constant that we are looking for, see [71]. These eigenvalues can be regarded as the eigenvalues of the Dirichlet to Neumann map for the operator −
+ I acting in the space H 1 (Ω). At this point, we have to mention the results of Escobar [57–59] for the Steklov eigenvalue problem



ϕ = 0 ∂ϕ ∂ν = λϕ

in Ω, on ∂Ω.

This problem was introduced by Steklov [115] and was initially studied by Calderon [25], because the set of eigenvalues for this problem coincides with the eigenvalues for the Dirichlet to Neumann map. Escobar proves some estimates and isoperimetric results for this problem on manifolds. Going back to the Sobolev trace embedding in full generality, one can consider W 1,p (Ω) → Lq (∂Ω), where W 1,p (Ω) is the usual Sobolev space of functions u ∈ Lp (Ω) with ∇u ∈ (Lp (Ω))N endowed with the norm uW 1,p (Ω) =



p

Ω

|∇u| dx +



p

Ω

|u| dx

1/p

.

Then, problems with nonlinear boundary conditions also appear in a natural way when one considers the Sobolev trace inequality in W 1,p (Ω) and Lq (∂Ω), p

p

SuLq (∂Ω)  uW 1,p (Ω) ,

1  q  p∗ =

p(N − 1) . N −p

318

J.D. Rossi

The best constant S is given by S(Ω, p, q) =

inf

1,p

u∈W 1,p (Ω)\W0 (Ω)



p Ω|∇u|

+ |u|p dx . ( ∂Ω |u|q dσ )p/q

The extremals (if there exist) are weak solutions of 
p u = |u|p−2 u in Ω, q−2 u |∇u|p−2 ∂u ∂ν = λ|u|

on ∂Ω.

(2.5)

Here,
p u = div(|∇u|p−2 ∇u) is the well-known operator called the p-Laplacian. Remark that (2.5) is a quasilinear elliptic problem. Also in this case we find an eigenvalue problem. When q = p, problem (2.5) becomes 
p u = |u|p−2 u in Ω, p−2 u on ∂Ω. |∇u|p−2 ∂u ∂ν = λ|u|

This is a nonlinear eigenvalue problem. Also in this case it can be proved that there exists a sequence of variational eigenvalues λn → ∞, see [71]. As for p = 2, the first eigenvalue, which is isolated and simple (see [100]), is related to the best constant for the Sobolev trace embedding. However, as happens for the eigenvalue problem for the p-Laplacian with Dirichlet boundary conditions, it is not known if the obtained sequence constitutes the whole spectrum. Now, for the sake of completeness, we will provide an answer for the question: Among the functions f : ∂Ω → R, which are the trace of a function of W 1,p (Ω)? To answer this question we have to introduce the fractional Sobolev spaces. We follow the presentation made by Bourgain, Brezis and Mironescu in [21]. D EFINITION 2.1. For 0 < s < 1, we define


  |u(x) − u(y)|p p s,p p W (A) = u ∈ L (A): [u]s,p := dx dy < ∞ N +sp A A |x − y| with the norm uW s,p (A) = uLp (A) + (p(1 − s))1/p [u]s,p . The quantity [·]s,p is known as the Gagliardo seminorm. The factor (p(1 − s))1/p that appears in the definition of the norm guarantees that lims→1 uW s,p (A) = uW 1,p (A) , see [21]. 1,p Let us consider the quotient space W 1,p (Ω)/W0 (Ω) with the quotient norm u = infu|∂Ω =v|∂Ω vW 1,p (Ω) . We have the following theorem. T HEOREM 2.1. For 1 < p < ∞, the trace operator T:

W 1,p (Ω) 1,p

W0 (Ω)

→ W 1−1/p,p (∂Ω)

Nonlinear boundary conditions

319

is a linear homeomorphism.  Some remarks on regularity. Now we want to state some regularity results for elliptic problems with nonlinear boundary conditions. We are not going to provide the more general result nor enter into the details of the proofs. However we give a sketch of the arguments just to look at the ideas involved. Also we have to mention here that we are not facing any regularity constraints coming from domain regularity, since we are assuming that Ω is a smooth domain. See [51] for regularity results in polyhedra. Let us deal with a solution of  −
u + u = |u|p−2 u in Ω, (2.6) ∂u q−2 u on ∂Ω, ∂ν = λ|u|  estimates for the solutions with p and q critical or subcritical. Now we prove C α (Ω) of (2.6), see [81]. We include some details for the sake of completeness. T HEOREM 2.2. Every weak solution of (2.6) with 1  q  2(N − 1)/(N − 2), 1  p   2N/(N − 2) belongs to C α (Ω). First, we deal with the subcritical case. Namely, 1 < q < 2(N − 1)/(N − 2), 1 < p < 2N/(N − 2). The idea is to adapt the classical bootstrapping argument, taking into account the nonlinear boundary condition. We start by recalling some linear results. P ROPOSITION 2.1. (I) Assume that g ∈ Lr (Ω) with r > 2N/(N + 2) and let φ ∈ H 1 (Ω) be the weak solution to
−
φ + φ = g in Ω, (2.7) ∂φ on ∂Ω, ∂ν = 0 then φW 1,β (Ω)  CgLr (Ω) with β = N r/(N − r) > 2. (II) Assume that h ∈ Ls (∂Ω) with s > 2(N − 1)/N , and let ψ be the weak solution to problem


−
ψ + ψ = 0 in Ω, ∂ψ ∂ν

=h

on ∂Ω.

(2.8)

Then ψW 1,γ (Ω)  ChLs (∂Ω) with γ = N s/(N − 1) > 2. P ROOF. Part (I) can be considered as the simplest case of the results in [114]. In this case the proof is easier: just integrating by parts we find       ∇φ∇ρ dx + φρ dx   gLr (Ω) ρLr ′ (Ω) ,  Ω

Ω

320

J.D. Rossi ′

with 1r + r1′ = 1, and by Sobolev embedding we can take a test function ρ ∈ W 1,β (Ω) with β ′ = N r ′ /(N + r ′ ). As a consequence, using Proposition 1 of [29] we get φ ∈ W 1,β (Ω) and φW 1,β (Ω)  CgLr (Ω) , where β = N r/(N − r), and since r > 2N/(N + 2), it follows β > 2. As for part (II), if ψ is the weak solution, multiplying by a regular test function η ∈ C 1 (Ω) we get        hLs (∂Ω) η s ′  ∇ψ∇η dx + ψη dx L (∂Ω) ,   Ω

Ω

where 1s + s1′ = 1. Then by density we can take η ∈ W 1,γ (Ω) and therefore, by the ′ ′ ′ ′ trace theorem, η|∂Ω ∈ W 1−(1/γ ),γ (∂Ω) ⊂ Lγ (N −1)/(N −γ ) (∂Ω), where s ′ = γ ′ (N − 1)/ (N − γ ′ ), which implies that γ = N s/(N − 1). Hence, by Proposition 1 of [29], we get that ψ ∈ W 1,γ (Ω) and ψW 1,γ (Ω)  ChLs (∂Ω) . This finishes the proof.  P ROOF OF T HEOREM 2.2. We decompose our original problem, taking g = |u|p−2 u and h = λ|u|q−2 u, in such a way that u = φ + ψ, where φ and ψ are the corresponding solutions to the linear problems (2.7) and (2.8). The idea to prove regularity for solutions of (2.6) is that we can iterate the estimates in Proposition 2.1, improving from step to step the regularity of u. The argument is as follows: We start assuming g ∈ Lr0 (Ω) and h ∈ Ls0 (∂Ω), where r0 =

2N (N − 2)(p − 1)

and s0 =

2(N − 1) . (N − 2)(q − 1)

In particular, if r0 > N/2 (that is, p < (N + 2)/(N − 2)) we get an exponent β0 > N  On the other hand, if q < N/(N − 2), we get that such that φ ∈ W 1,β0 (Ω) ⊂ C α (Ω).  As a consequence, the C α reguψ ∈ W 1,γ0 (Ω) with γ0 > N and in this case ψ ∈ C α (Ω). larity for u is proved in the case q < N/(N − 2), p < (N + 2)/(N − 2). If not, in any case we have proved that u ∈ W 1,τ0 (Ω) with τ0 = min{β0 , γ0 } > 2. Then we can iterate exactly the same calculation as before, starting with g ∈ Lr1 (Ω) and h ∈ Ls1 (∂Ω), where r1 =

N τ0 (N − τ0 )(p − 1)

and s1 =

(N − 1)τ0 . (N − τ0 )(q − 1)

If r1 and s1 were both large enough (namely, r1 > N/2 and s1 > N − 1), then we have finished. If not, we get that u ∈ W 1,τ1 (Ω), where τ1 =



N s1 N −1

 s1 N r1 min NN−1 , N −r1

if r1 > if r1 

N 2 and s1 N 2.

 N − 1,

Nonlinear boundary conditions

321

Let us estimate these quantities in terms of the starting exponent τ0 . Since τ0 > 2, we have N s1 N N = τ0
τ0 . N − 1 (N − τ0 )(q − 1) (N − 2)(q − 1) And, on the other hand, it is easy to see that N N N r1 = τ0 > τ0 . N − r1 p(N − τ0 ) − N p(N − 2) − N Therefore (taking into account that p and q are subcritical) we have proved that there exists a constant C = C(N, p, q) > 1 such that τ1
Cτ0 , and, in general, τk
C k τ0 . This ∗ implies that in a finite number of steps we reach that u ∈ W 1,τ (Ω) with τ ∗ > N , and hence α  u ∈ C (Ω). Next, we will sketch briefly the arguments in the critical case, p = 2∗ . In this case, the problem comes from the first iteration, since there is no margin to improve directly the initial exponent, getting W 1,β0 (Ω) regularity for some β0 > 2. To overcome this difficulty we can use a truncation argument by Trudinger (see [120]) which proves that uLτ (Ω)  C(|Ω|, uLp∗ (Ω) ), where τ > p ∗ . The sketch of the argument is as follows: consider the problem 

∗ −2

−
u + u = λ|u|2 ∂u q−2 u ∂ν = |u|

u

in Ω, on ∂Ω,

where q is subcritical. Assume that u ∈ H 1 (Ω), u > 0, is a solution and let us prove that u ∈ Lτ (Ω) for some τ > 2∗ . The main idea is to choose a suitable truncation of uβ as test function with β greater but close to one. After some manipulations, that in our case involve the Sobolev trace inequality to handle the integrals over the boundary that appear, ∗ we arrive to u ∈ Lβ2 (Ω). As β is greater than one this estimate gives the required starting  point, after which the argument follows as in the previous case, getting finally u ∈ C α (Ω). The case q = 2∗ with p subcritical can be handled in a similar way. With the argument given by Trudinger [120], we can begin the iterative procedure and also in this case we get  u ∈ C α (Ω).  From the results of Cherrier [29], we have that weak positive solutions of (2.6)  See also [4], [84] and [119] for regularity results. Also one can extend the are C ∞ (Ω).  C α (Ω)-regularity of the solutions using arguments based in [114] for more general elliptic problems like 

 − div a(x)∇v + v = h

∂v a(x) ∂ν =g

See [14] for the details.

in Ω, on ∂Ω.

322

J.D. Rossi

3. Existence results for an elliptic problem with nonlinear boundary conditions. A variational approach To begin the study of existence of solutions to problems with nonlinear boundary conditions let us see how the usual variational techniques can be adapted. First, we follow the ideas from J. Fernández Bonder and the author, see [71]. We study the existence of nontrivial solutions for the following problem 


p u = |u|p−2 u

|∇u|p−2 ∂u ∂ν

= f (u)

in Ω,

(3.1)

on ∂Ω.

As we explained in the Introduction (see also Section 2), problems of the form (3.1) appears in a natural way when one considers the Sobolev trace inequality S 1/p uLq (∂Ω)  uW 1,p (Ω) ,

1  q  p∗ =

p(N − 1) . N −p

In fact, the extremals (if there exists) are solutions of (3.1) for f (u) = λ|u|q−2 u. For weak solutions of (3.1) we understand critical points of the associated energy functional   1 p p |∇u| + |u| dx − F (u) dσ, (3.2) F(u) = p Ω ∂Ω where F ′ (u) = f (u). We fix 1 < p < N and look for conditions on the nonlinear term f (u) that provide us with the existence of nontrivial solutions of (3.1). This functional F is well defined and C 1 in W 1,p (Ω) if f has a critical or subcritical growth, namely |f (u)|  C(1 + |u|q ) with 1  q  p∗ = p(N − 1)/(N − p). Moreover, in the subcritical case 1 < q < p ∗ , the immersion W 1,p (Ω) ֒→ Lq (∂Ω) is compact while in the critical case q = p∗ is only continuous, see Section 2. First, we deal with a superlinear and subcritical nonlinearity. For simplicity we will consider f (u) = λ|u|q−2 u,

(3.3)

where q verifies 1 < q < p∗ = p(N − 1)/(N − p). We prove the following theorems using standard variational arguments together with the Sobolev trace immersion that provide the necessary compactness. See [79] for similar results for the p-Laplacian with Dirichlet boundary conditions. We divide the presentation in three cases according to p < q, p = q and p > q. First, for p < q < p∗ we have the following theorem. T HEOREM 3.1. Let f satisfy (3.3) with p < q < p∗ , then there exists infinitely many nontrivial solutions of (3.1) which are unbounded in W 1,p (Ω).

Nonlinear boundary conditions

323

Now we proceed with a sketch of the proof of Theorem 3.1. We believe that the ideas involved are illustrative on how to deal with nonlinear boundary conditions from a variational point of view. Let us begin with the following lemma that will be helpful in order to prove the Palais– Smale condition for the functional (3.2). L EMMA 3.1. Let φ ∈ W 1,p (Ω)′ . Then there exists a unique weak solution u ∈ W 1,p (Ω) of −
p u + |u|p−2 u = φ.

(3.4)

Moreover, the operator Ap : φ → u is continuous. P ROOF. Let us observe that weak solutions u ∈ W 1,p (Ω) of (3.4) are critical points of the functional  1 I (u) = |∇u|p + |u|p dx − φ, u, p Ω where ·, · denotes the duality paring in W 1,p (Ω). Hence, existence and uniqueness are a consequence of the fact that I is a weakly lower semicontinuous, strictly convex and bounded below functional. For the continuous dependence, let us first recall the following inequality (cf. [113])

p−2  |x| x − |y|p−2 y, x − y




Cp |x − y|p

|x−y|2 Cp (|x|+|y|) 2−p

if p
2, if p  2,

(3.5)

where (·, ·) denotes the usual scalar product in Rm . Now, given φ1 , φ2 ∈ W 1,p (Ω)′ let us consider u1 , u2 ∈ W 1,p (Ω) the corresponding solutions of problem (3.4). Then, for i = 1, 2 we have 

Ω

|∇ui |p−2 ∇ui (∇u1 − ∇u2 ) dx +



Ω

|ui |p−2 ui (u1 − u2 ) − φi (u1 − u2 ) dx = 0.

Hence, substracting and using inequality (3.5) we obtain, for p
2, Cp



Ω

|∇u1 − ∇u2 |p + |u1 − u2 |p dx

! "  (φ1 − φ2 ), (u1 − u2 )

 φ1 − φ2 W 1,p (Ω)′ u1 − u2 W 1,p (Ω) .

324

J.D. Rossi

Therefore, Ap (φ1 ) − Ap (φ2 )W 1,p (Ω)  C(φ1 − φ2 W 1,p (Ω)′ )1/(p−1) . Now, for the case p  2, we first observe that    ∇(u1 − u2 )p dx Ω





Ω

|∇(u1 − u2 )|2 dx (|∇u1 | + |∇u2 |)2−p

p/2 

Ω

p |∇u1 | + |∇u2 | dx

(2−p)/2

and 

Ω

|u1 − u2 |p dx





Ω

|u1 − u2 |2 dx (|u1 | + |u2 |)2−p

p/2 

Ω

p |u1 | + |u2 | dx

(2−p)/2

.

As in the previous case, we get u1 − u2 W 1,p (Ω)

(u1 W 1,p (Ω) + u2 W 1,p (Ω) )2−p

 Cφ1 − φ2 W 1,p (Ω)′ .

(3.6)

p

Now we observe that ui W 1,p (Ω)  φi W 1,p (Ω)′ ui W 1,p (Ω) . Hence, (3.6) becomes   Ap (φ1 ) − Ap (φ2 ) 1,p W (Ω)

1/(p−1) 1/(p−1) 2−p  C φ1 W 1,p (Ω)′ + φ2 W 1,p (Ω)′ φ1 − φ2 W 1,p (Ω)′

and the proof is finished.



With this lemma we can verify the Palais–Smale condition for F . L EMMA 3.2. The functional F satisfies the Palais–Smale condition. P ROOF. Let (uk )k
1 ⊂ W 1,p (Ω) be a Palais–Smale sequence, that is a sequence such that F(uk ) → c

and F ′ (uk ) → 0.

(3.7)

Let us first prove that (3.7) implies that (uk ) is bounded. From (3.7) it follows that there exists a sequence εk → 0 such that |F ′ (uk )w|  εk wW 1,p (Ω) for all w ∈ W 1,p (Ω). Now we have 1 ′ 1 F (uk )uk + F ′ (uk )uk q q  1 1 1 p uk W 1,p (Ω) + F ′ (uk )uk − = p q q

c + 1
F(uk ) −

Nonlinear boundary conditions

1 1 p uk W 1,p (Ω) − − p q  1 1 p uk W 1,p (Ω) − −
p q




325

1 uk W 1,p (Ω) εk q 1 uk W 1,p (Ω) . q

Hence, uk is bounded in W 1,p (Ω). By compactness we can assume that uk ⇀ u weakly in W 1,p (Ω) and uk → u strongly in Lq (∂Ω) and a.e. in ∂Ω. Then, as p < q < p∗ , it follows ′ that |uk |q−2 uk → |u|q−2 u in Lp∗ (∂Ω) and hence in W 1,p (Ω)′ . Therefore, according to q−2 Lemma 3.1, uk → Ap (|u| u) in W 1,p (Ω). This completes the proof.  Now we introduce a topological tool, the genus, see [91]. D EFINITION 3.1. Given a Banach space X, we consider the class Σ = {A ⊂ X: A is closed, A = −A}. Over this class we define the genus, γ : Σ → N ∪ {∞}, as 

 γ (A) = min k ∈ N: there exists ϕ ∈ C A, Rk \ {0} , ϕ(x) = −ϕ(−x) .

We will use the following proposition whose proof can be found in [6].

P ROPOSITION 3.1 ([6], Theorem 2.23). Let F : X → R verifying: (1) F ∈ C 1 (X) and even. (2) F verifies the Palais–Smale condition. (3) There exists a constant r > 0 such that F(u) > 0 in 0 < uX < r and F(u)
c > 0 if uX = r. (4) There exists a closed subspace Em ⊂ X of dimension m, and a compact set Am ⊂ Em such that F < 0 on Am and 0 lies in a bounded component of Em − Am in Em . Let B be the unit ball in X, we define

and

Then



 Γ = h ∈ C(X, X): h(0) = 0, h is an odd homeomorphism and F h(B)
0  Km = K ⊂ X: K = −K, K is compact,

 and γ K ∩ h(∂B)
m for all h ∈ Γ . cm = inf max F(u) K∈Km u∈K

is a critical value of F , with 0 < c  cm  cm+1 < ∞. Moreover, if cm = cm+1 = · · · = cm+r then γ (Kcm )
r + 1 where Kcm = {u ∈ X: F ′ (u) = 0, F(u) = cm }.

326

J.D. Rossi

P ROOF OF T HEOREM 3.1. We need to check the hypotheses of Proposition 3.1. The fact that F is C 1 is a straightforward adaptation of the results in [110]. The Palais–Smale condition was already checked in Lemma 3.2. Let us check (3). From the Sobolev trace theorem, we obtain 1 λ 1 λ p q p q uW 1,p (Ω) − uLq (∂Ω)
uW 1,p (Ω) − C uW 1,p (Ω) p q p q 

= g uW 1,p (Ω) ,

F(u) =

where g(t) = p1 t p − C qλ t q . As q > p, (3) follows for r = r(C, λ, p, q) small. Finally, to verify (4), let us consider a sequence of subspaces Em ⊂ W 1,p (Ω) of dimension m such that Em ⊂ Em+1 and u|∂Ω ≡ 0 for u = 0, u ∈ Em . Hence, min



u∈Bm ∂Ω

|u|q dσ > 0,

where Bm = {u ∈ Em : uW 1,p (Ω) = 1}. Now we observe that λt q tp min F(tu)  uW 1,p (Ω) − p q u∈Bm



∂Ω

|u|q dσ < 0

for all u ∈ Bm and t
t0 . Therefore, (4) follows by taking Am = t0 Bm .



In order to see that the sequence of critical points of F that we have found is unbounded in W 1,p (Ω), we need the following result. L EMMA 3.3. Let (cm ) ⊂ R be the sequence of critical values given by Theorem 3.1. Then limm→∞ cm = ∞. p

q

1 P ROOF. Let M = {u ∈ W 1,p (Ω) \ {0}: λp uW 1,p (Ω)  uLq (∂Ω) }. By the Sobolev trace theorem, there exists a constant r > 0 such that q

r < uLq (∂Ω)

∀u ∈ M.

(3.8)

Let us define bm = sup h∈Γ

inf

c {u∈∂B∩Em−1 }

 F h(u) .

It is proved in [6] that bm  cm , hence to prove our result it is enough to show that bm → ∞. Now, bm+1
infu∈∂B∩Emc F(h(u)) for all h ∈ Γ . We will construct h˜ m ∈ Γ such that lim

inf

c m→∞ u∈∂B∩Em

 F h˜ m (u) = ∞.

Nonlinear boundary conditions

327

c } and observe First, let us define the following sequence dm = inf{uW 1,p (Ω) : u ∈ M ∩ Em c that dm → ∞. In fact, if not, there exists a sequence um ∈ M ∩ Em such that um ⇀ 0 weakly in W 1,p (Ω) and therefore um → 0 in Lq (∂Ω), a contradiction with (3.8). Next, let us consider hm (u) = R −1 dm u where R > 1 is to be fixed. From hm we will construct h˜ m . Given u ∈ W 1,p (Ω) such that u|∂Ω ≡ 0, pick β = β(u) such that

1 p q βuW 1,p (Ω) = βuLq (∂Ω) , λp so βu ∈ M. If we consider g(t) = F(tu) with u|∂Ω ≡ 0, it is easy to see that g is increasing c ∩B in [0, β(u)] so g achieves its maximum on that interval for t = β(u). Take u0 ∈ Em such that u0 |∂Ω ≡ 0, then for R > 1, R −1 dm  dm  βu0 W 1,p (Ω) = β(u0 ). c ∩ B such that This inequality implies that, for every R > 1 and for every u0 ∈ Em −1 u0 |∂Ω ≡ 0, it holds F(hm (u0 )) = F(R dm u0 )
0. As hm (0) = 0, it follows that c ∩ B) ⊂ {u ∈ W 1,p (Ω): F(u)
0}. Therefore, h | c satisfies the requirements hm (Em m Em needed in order to belong to Γ so it comes natural try to extend hm to W 1,p (Ω) so it c ∩ B) + ε(E ∩ B). Let us see that belongs to Γ . Given ε > 0, consider Zε = dm R −1 (Em m c for ε small, Zε ⊂ M . If not, there exists a sequence εj → 0 and a sequence (uj ) ⊂ M such that uj ∈ Zεj . In particular, uj is bounded in W 1,p (Ω) so we can assume that

uj ⇀ u

weakly in W 1,p (Ω),

uj → u

in Lq (∂Ω).

Moreover, as uj ∈ M it follows that u|∂Ω ≡ 0. On the other hand, as  · W 1,p (Ω) is weakly c ∩ B), a conlower semicontinuous, we have that u ∈ M and, as εj → 0, u ∈ dm R −1 (Em c tradiction. So we have proved that there exists ε0 > 0 such that Zε0 ⊂ M . This fact allows us to define
−1 ˜hm (u) = hm (u) = dm R u ε0 u

c, if u ∈ Em if u ∈ Em .

Now, if u ∈ Em ∩ B we have h˜ m (u) = ε0 u ∈ Zε0 ⊂ M c , then

 1 λ p q F h˜ m (u) = F(ε0 u) = ε0 uW 1,p (Ω) − ε0 uLq (∂Ω) p q   λ q −1 1 p p q ε0 uW 1,p (Ω) + ε0 uW 1,p (Ω) − ε0 uLq (∂Ω) = q λp λp
0,

328

J.D. Rossi

c and u ∈ E ∩ B, we that is, given u ∈ B if we decompose u = u1 + u2 with u1 ∈ Em 2 m obtain h˜ m (u) = h˜ m (u1 ) + h˜ m (u2 ) = dm R −1 u1 + ε0 u2 ∈ Zε0 ⊂ M c from where it follows that F(h˜ m (u))
0 and hence h˜ m ∈ Γ . c , but this Finally, we need to prove that F(h˜ m (u)) → ∞ as m → ∞ for u ∈ ∂B ∩ Em c follows from the facts that dm → ∞, that dm  β(u) for u ∈ B ∩ Em and that we can c,h ˜ m (u) = dm R −1 u and choose R large enough. If u ∈ ∂B ∩ Em

 (dm R −1 )p λ(dm R −1 )q p q uW 1,p (Ω) − uLq (∂Ω) F h˜ m (u) = p q 

  λ q −1 p 1 −1 q−p = dm R − dm R uLq (∂Ω) p q 

  λ q −1 p 1 −1 q−p
dm R − β(u)R uLq (∂Ω) p q 

 R p−q −1 p 1 . = dm R − p pq

As q > p we conclude that if R is large enough, then F(h˜ m (u)) → +∞.



Now we consider the case 1 < q < p. Using the genus and that the functional F verifies a Palais–Smale condition, we have, T HEOREM 3.2. Let f satisfy (3.3) with 1 < q < p, then there exists infinitely many nontrivial solutions of (3.1) which form a compact set in W 1,p (Ω). The proof of Theorem 3.2 follows from a series of lemmas, the proofs will be omitted or sketched, see [71] for details. L EMMA 3.4. For every n ∈ N there exists a constant ε > 0 such that

 γ F −ε
n,

where F c = {u ∈ W 1,p (Ω): F(u)  c}.

L EMMA 3.5. The functional F is bounded below and verifies the Palais–Smale condition. The following two propositions give us the proof of Theorem 3.2. P ROPOSITION 3.2. Let  Σk = A ⊂ W 1,p (Ω) \ {0}: A is closed, A = −A, and γ (A)
k ,

where γ stands for the genus. Then ck = inf sup F(u) A∈Σk u∈A

Nonlinear boundary conditions

329

is a negative critical value of F and moreover, if c = ck = · · · = ck+r , then γ (Kc )
r + 1, where Kc = {u ∈ W 1,p (Ω): F(u) = c, F ′ (u) = 0}. P ROOF. According to Lemma 3.4 for every k ∈ N there exists ε > 0 such that γ (F −ε )
k. As F is even and continuous it follows that F −ε ∈ Σk therefore ck  −ε < 0. Moreover, by Lemma 3.5, F is bounded below so ck > −∞. One can see that ck is in fact a critical value for F . To this end let us suppose that c = ck = · · · = ck+r . As F is even it follows that Kc is symmetric. The Palais–Smale condition implies that Kc is compact, therefore if γ (Kc )  r by the continuity property of the genus (see [110]) there exists a neighborhood of Kc , Nδ (Kc ) = {v ∈ W 1,p (Ω): d(v, Kc )  δ} such that γ (Nδ (Kc )) = γ (Kc )  r. By the usual deformation argument, we get η(1, F c+ε/2 − Nδ (Kc )) ⊂ F c−ε/2 . On the other hand, by the definition of ck+r , there exists A ⊂ Σk+r such that A ⊂ F c+ε/2 , hence

 η 1, A − Nδ (Kc ) ⊂ F c−ε/2 .

(3.9)

Now by the monotonicity of the genus (see [110]), we have γ (A − Nδ (Kc ))
γ (A) − γ (Nδ (Kc ))
k. As η(1, ·) is an odd homeomorphism, it follows that (see [110]) γ (η(1, A − Nδ (Kc )))
γ (A − Nδ (Kc ))
k. But as η(1, A − Nδ (Kc )) ∈ Σk then sup u∈η(1,A−Nδ (Kc ) )

F(u)
c = ck , 

a contradiction with (3.9). Now we show that the critical points of F are a compact set of W 1,p (Ω). P ROPOSITION 3.3. The set K = {u ∈ W 1,p (Ω): F ′ (u) = 0} is compact in W 1,p (Ω).

P ROOF. As F is C 1 it is immediate that K is closed. Let uj be a sequence in K. We have that  p 0 = F ′ (uj )uj = uj W 1,p (Ω) − λ |uj |q dσ ∂Ω




p uj W 1,p (Ω)

q

− Cλuj W 1,p (Ω) .

As 1 < q < p, we conclude that uj is bounded in W 1,p (Ω). Now we can use Palais–Smale condition to extract a convergent subsequence.  In the case p = q, the equation and the boundary condition are homogeneous of the same degree, so we are dealing with a nonlinear eigenvalue problem. In the linear case, that is, for p = 2, this eigenvalue problem is known as the Steklov problem [115]. We have the following result whose proof can be found in [71], we do not provide the details in this case.

330

J.D. Rossi

T HEOREM 3.3. Let f satisfy (3.3) with p = q, then there exists a sequence of eigenvalues λn of (3.1) such that λn → +∞ as n → +∞. The variational eigenvalues λk can be characterized by p

uLp (∂Ω) 1 , = sup min λk C∈Ck u∈C up 1,p W

(3.10)

(Ω)

where Ck = {C ⊂ W 1,p (Ω); C is compact, symmetric and γ (C)
k} and γ is the genus (see [91] and Definition 3.1). It is shown in [73] that there exists a second eigenvalue for (6.4) and that it coincides with the second variational eigenvalue λ2 . Moreover, the following characterization of the second eigenvalue λ2 holds λ2 = inf




u∈A

Ω

|∇u| + |u| dx , p

p

where A = {u ∈ W 1,p (Ω); uLp (∂Ω) = 1 and |∂Ω ± | > 0}, with ∂Ω + = {x ∈ ∂Ω; u(x) > 0} and ∂Ω − defined analogously. Next we consider the critical growth on f . In this case the compactness of the immersion W 1,p (Ω) ֒→ Lp∗ (∂Ω) fails, so in order to recover some sort of compactness, in the same spirit of [22], we consider a perturbation of the critical power, that is, f (u) = |u|p∗ −2 u + λ|u|r−2 u = |u|p(N−1)/(N −p)−2 u + λ|u|r−2 u.

(3.11)

Here we use the concentration–compactness method introduced in [96,97] and follow ideas from [80]. First, we have the following theorem. T HEOREM 3.4. Let f satisfy (3.11) with p < r < p∗ , then there exists a constant λ0 > 0 depending on p, r, N and Ω, such that if λ > λ0 , problem (3.1) has at least a nontrivial solution in W 1,p (Ω). To prove this existence result, since we have lost the compactness in the inclusion W 1,p (Ω) ֒→ Lp∗ (∂Ω), we can no longer expect the Palais–Smale condition to hold. Anyway we can prove a local Palais–Smale condition that will hold for F(u) below a certain value of energy. Let uj be a bounded sequence in W 1,p (Ω) then there exists a subsequence that we still denote uj , such that uj ⇀ u

weakly in W 1,p (Ω),

strongly in Lr (∂Ω), 1  r < p∗ ,  p |∇uj |p ⇀ dµ, uj |∂Ω  ∗ ⇀ dη,

uj → u

weakly-* in the sense of measures. Observe that dη is a measure supported on ∂Ω.

Nonlinear boundary conditions

331

 from the Sobolev trace inequality we obtain, passing to the If we consider φ ∈ C ∞ (Ω), limit, 

∂Ω



|φ|p∗ dη



1/p∗

S 1/p

p

Ω

|φ| dµ +



Ω

p

p

|u| |∇φ| dx +



Ω

p

|φu| dx

1/p

,

(3.12)

where S is the best constant in the Sobolev trace embedding theorem. From (3.12) we observe that, if u = 0 we get a reverse Hölder-type inequality (but it involves one integral over ∂Ω and one over Ω) between the two measures µ and η. Now we state the following lemma due to Lions [96,97]. L EMMA 3.6. Let uj be a weakly convergent sequence in W 1,p (Ω) with weak limit u such that |∇uj |p ⇀ dµ

and

  uj |∂Ω p∗ ⇀ dη,

weakly-* in the sense of measures. Then there exists x1 , . . . , xl ∈ ∂Ω such that  (1) dη = |u|p∗ + lj =1 ηj δxj , ηj > 0,  (2) dµ
|∇u|p + lj =1 µj δxj , µj > 0, (3) (ηj )p/p∗  µj /S. Next, we use Lemma 3.6 to prove a local Palais–Smale condition. L EMMA 3.7. Let uj ⊂ W 1,p (Ω) be a Palais–Smale sequence for F , with energy level c. If c < ( p1 − p1∗ )S p∗ /(p∗ −p) , where S is the best constant in the Sobolev trace inequality, then there exists a subsequence ujk that converges strongly in W 1,p (Ω). P ROOF. From the fact that uj is a Palais–Smale sequence it follows that uj is bounded in W 1,p (Ω) (see Lemma 3.2). By Lemma 3.6 there exists a subsequence, that we still denote uj , such that uj ⇀ u uj → u

weakly in W 1,p (Ω), in Lr (∂Ω), 1 < r < p∗ , and a.e. in ∂Ω,

|∇uj |p ⇀ dµ
|∇u|p +

l 

µk δxk ,

k=1

l      uj |∂Ω p∗ ⇀ dη = u|∂Ω p∗ + ηk δxk . k=1

(3.13)

332

J.D. Rossi

Let φ ∈ C ∞ (RN ) such that φ ≡ 1 in B(xk , ε), φ ≡ 0 in B(xk , 2ε)c and |∇φ|  2ε , where xk belongs to the support of dη. Consider {uj φ}. Obviously this sequence is bounded in W 1,p (Ω). As F ′ (uj ) → 0 in W 1,p (Ω)′ , we obtain that F ′ (uj ); φuj  → 0 as j → ∞. By (3.13), we obtain 

lim

j →∞ Ω

=



|∇uj |p−2 ∇uj ∇φuj dx φ dη + λ

∂Ω



∂Ω

|u|r φ dσ −



φ dµ −

Ω



Ω

|u|p φ dx.

Now, by Hölder inequality and weak convergence, we obtain     p−2  0  lim  |∇uj | ∇uj ∇φuj dx  j →∞ Ω

 lim

j →∞

C





Ω

|∇uj |p dx

N

B(xk ,2ε)∩Ω

C

(p−1)/p 



B(xk ,2ε)∩Ω

|∇φ| dx

Ω

|∇φ|p |uj |p dx

1/N 

|u|pN/(N−p) dx

1/p

pN/(N−p)

|u|

B(xk ,2ε)∩Ω

(N −p)/(pN)

dx

(N −p)/(pN)

→ 0 as ε → 0.

Then lim

ε→0

'

∂Ω

φ dη + λ



∂Ω

r

|u| φ dσ −



Ω

φ dµ −



Ω

p

(

|u| φ dx = ηk − µk = 0. (3.14)

By Lemma 3.6 we have that (ηk )p/p∗ S  µk , therefore by (3.14) we get (ηk )p/p∗ S  ηk . Then, either ηk = 0 or ηk
S p∗ /(p∗ −p) .

(3.15)

If (3.15) does indeed occur for some k0 then, from the fact that uj is a Palais–Smale sequence, we obtain c = lim F(uj ) j →∞

" 1! ′ F (uj ); uj j →∞ p    1 1 1 1 p∗ S p∗ /(p∗ −p) − − |u| dσ +
p p∗ p p∗ ∂Ω

= lim F(uj ) −

Nonlinear boundary conditions

333

 1 1 − |u|r dσ p r ∂Ω  1 1 S p∗ /(p∗ −p) . −
p p∗ +λ



(3.16)

  As c < ( p1 − p1∗ )S p∗ /(p∗ −p) , it follows that ∂Ω |uj |p∗ dσ → ∂Ω |u|p∗ dσ and therefore  uj → u in Lp∗ (∂Ω). Now the proof finishes using the continuity of Ap . P ROOF OF T HEOREM 3.4. In view of the previous result, we seek for critical values below level c. For that purpose, we want to use the mountain pass lemma. Hence we have to check the following conditions: (1) There exist constants R, r > 0 such that if uW 1,p (Ω) = R, then F(u) > r. (2) There exists v0 ∈ W 1,p (Ω) such that v0 W 1,p (Ω) > R and F(v0 ) < r. Let us first check (1). By the Sobolev trace theorem, we have F(u) =

1 1 p uW 1,p (Ω) − p p∗



∂Ω

|u|p∗ dσ −

λ r



∂Ω

|u|r dσ

1 λ 1 p p
uW 1,p (Ω) − S p∗ uW∗1,p (Ω) − CurW 1,p (Ω) . p p∗ r Let g(t) =

1 p λ 1 t − S p∗ t p∗ − Ct r . p p∗ r

It is easy to check that g(R) > r for some R, r > 0. (2) is immediate as for a fixed w ∈ W 1,p (Ω) with w|∂Ω ≡ 0, we have lim F(tw) = −∞.

t→∞

Now, the candidate for critical value according to the mountain pass theorem is

 c = inf sup F φ(t) , φ∈C t∈[0,1]

where C = {φ : [0, 1] → W 1,p (Ω); continuous and φ(0) = 0, φ(1) = v0 }. The problem is to show that c < ( p1 − p1∗ )S p∗ /(p∗ −p) in order to apply the local Palais–Smale condition. We fix w ∈ W 1,p (Ω) with wLp∗ (∂Ω) = 1, and define h(t) = F(tw). We want to study the maximum of h. As limt→∞ h(t) = −∞ it follows that there exists a tλ > 0 such that supt>0 F(tw) = h(tλ ). Differentiating we obtain p−1

0 = h′ (tλ ) = tλ

p

p −1

wW 1,p (Ω) − tλ ∗

− tλr−1 λwrLr (∂Ω) ,

(3.17)

334

J.D. Rossi p

p −p

from where it follows that wW 1,p (Ω) = tλ ∗ p/(p −p)

p −r

wW 1,p∗(Ω) . From (3.17), as tλ ∗

r−p

+ tλ

λwrLr (∂Ω) . Hence tλ 

+ λwrLr (∂Ω) → ∞ when λ → ∞, we obtain that

lim tλ = 0.

λ→∞

(3.18)

On the other hand, it is easy to check that if λ > λ˜ it must be F(tλ˜ w)
F(tλ w), so by (3.18) we get limλ→∞ F(tλ w) = 0. But this identity means that there exists a constant λ0 > 0 such that if λ
λ0 , then  1 1 sup F(tw) < S p∗ /(p∗ −p) , − p p∗ t
0 and the proof is finished if we choose v0 = t0 w with t0 large in order to have F(t0 w) < 0.  Now we prove a second result for the critical case. T HEOREM 3.5. Let f satisfy (3.11) with 1 < r < p, then there exists a constant λ1 > 0 depending on p, r, N and Ω such that if 0 < λ < λ1 , problem (3.1) has infinitely many nontrivial solutions in W 1,p (Ω). We begin, as we have done previously, proving a local Palais–Smale condition using Lemma 3.6. L EMMA 3.8. Let (uj ) ⊂ W 1,p (Ω) be a Palais–Smale sequence for F , with energy level c. If c < ( p1 − p1∗ )S p∗ /(p∗ −p) − Kλp∗ /(p∗ −r) , where K depends only on p, r, N , and |∂Ω|, then there exists a subsequence (ujk ) that converges strongly in W 1,p (Ω). P ROOF. From the fact that uj is a Palais–Smale sequence it follows that uj is bounded in W 1,p (Ω) (see Lemmas 3.2 and 3.7). Now the proof follows exactly as in Lemma 3.7 until we get to    1 1 1 1 S p∗ /(p∗ −p) − − c
|u|p∗ dσ + p p∗ p p∗ ∂Ω   1 1 − +λ |u|r dσ, p r ∂Ω where u is the weak limit of uj in W 1,p (Ω). Applying Hölder inequality we find 

 1 1 1 1 p p∗ /(p∗ −p) c
S uL∗p∗ (∂Ω) − − + p p∗ p p∗  1 1 |∂Ω|1−r/p∗ urLp∗ (∂Ω) . − +λ p r

Nonlinear boundary conditions

335

Now, let f (x) = c1 x p∗ − λc2 x r . This function reaches its absolute minimum at x0 = ( pλc∗2cr1 )1/(p∗ −r) , that is, f (x)
f (x0 ) = −Kλp∗ /(p∗ −r) , where K = K(p, q, N, |∂Ω|). Hence c
( p1 − fore,  lim

1 p∗ /(p∗ −p) p∗ )S

j →∞ ∂Ω

|uj |p∗ dσ =

− Kλp∗ /(p∗ −r) , which contradicts our hypothesis. There

∂Ω

|u|p∗ dσ,

and the rest of the proof is as that of Lemma 3.7.



We observe, using the Sobolev trace theorem, that F(u)




1 p p uW 1,p (Ω) − c1 uW∗1,p (Ω) − λc2 urW 1,p (Ω) = j uW 1,p (Ω) , p

where j (x) = p1 x p − c1 x p∗ − λc2 x r . As j attains a local but not a global minimum (j is not bounded below), we have to perform some sort of truncation. To this end, let x0 , x1 be such that m < x0 < M < x1 where m is the local minimum of j and M is the local maximum and j (x1 ) > j (m). For these values x0 and x1 we can choose a smooth function τ (x) such that τ (x) = 1 if x  x0 , τ (x) = 0 if x
x1 and 0  τ (x)  1. Finally, let ϕ(u) = τ (uW 1,p (Ω) ) and define the truncated functional as follows    1 λ  =1 F(u) |∇u|p + |u|p dx − |u|r dσ. |u|p∗ ϕ(u) dσ − p Ω p∗ ∂Ω r ∂Ω 
j˜(uW 1,p (Ω) ), where j˜(x) = 1 x p − c1 x p∗ τ (x) − λc2 x r . We observe As above, F(u) p that if x  x0 then j˜(x) = j (x) and if x
x1 then j˜(x) = p1 x p − λc2 x r . . Now we state a lemma that contains the main properties of F

 is C 1 , if F(u)   0 then uW 1,p (Ω) < x0 and F(v) = F(v)  for every v L EMMA 3.9. F  satisfies a close enough to u. Moreover, there exists λ1 > 0 such that, if 0 < λ < λ1 then F local Palais–Smale condition for c  0.

P ROOF. We only have to check the local Palais–Smale condition. Observe that every  with energy level c  0 must be bounded, therefore by Palais–Smale sequence for F Lemma 3.8, if λ verifies 0 < ( p1 − p1∗ )S p∗ /(p∗ −p) − Kλp∗ /(p∗ −r) then there exists a convergent subsequence.  −ε )
n, where L EMMA 3.10. For every n ∈ N there exists ε > 0 such that γ (F −ε = {u, F(u)   −ε}. F P ROOF. The proof is analogous to that of Lemma 3.4.



P ROOF OF T HEOREM 3.5. The proof is analogous to that of Theorem 3.2, here we use Lemmas 3.8 and 3.10 instead of Lemmas 3.5 and 3.4, respectively, to work with the func and Lemma 3.9 to conclude on F . tional F 

336

J.D. Rossi

Next, we deal with supercritical growth on f . More precisely, we study a subcritical perturbation of the supercritical power, that is, we consider f (u) = λ|u|q−2 u + |u|r−2 u,

(3.19)

with q
p∗ > r > p. In this case, not only the compactness fails but also the functional F is not well defined in W 1,p (Ω), so we have to perform a truncation in the nonlinear term λ|u|q−2 u following ideas from [27]. For this case we have the following theorem. T HEOREM 3.6. Let f satisfy (3.19) with q
p∗ > r > p, then there exists a constant λ2 depending on p, q, r, N and Ω such that if 0 < λ < λ2 , problem (3.1) has a nontrivial positive solution in W 1,p (Ω) ∩ L∞ (∂Ω). P ROOF OF T HEOREM 3.6. Let us consider the following truncation of |u|q−2 u   0, h(u) = uq−1 ,  q−r r−1 K u ,

u < 0, 0  u < K, u
K.

Then h verifies h(u)  K q−r ur−1 . So we consider the truncated problem 


p u = up−1 |∇u|p−2 ∂u ∂ν

in Ω,

= λh(u) + ur−1

(3.20)

on ∂Ω,

and we look a positive nontrivial solution of (3.20) that satisfies u  K. Such a solution will be a nontrivial positive solution of (3.1). To this end, we consider the truncated functional Fλ (u) =

1 p



Ω

|∇u|p + |u|p dx − λ



∂Ω

H (u) dσ −



|u|r dσ, r

∂Ω

(3.21)

where H (u) verifies H ′ (u) = h(u). One can check that here exists a mountain pass solution u = uλ for (3.20), that is, a critical point of Fλ with energy level cλ . One can easily check that this least energy solution u is positive. Moreover, the energy level cλ is a decreasing function of λ, so we have that Fλ (u) = cλ  c0 . Now, using (3.21), (3.20) and that H (u)  1r h(u)u we have that c0
Fλ (u) =

1 p



Ω

|∇u|p + |u|p dx − λ



∂Ω

   1 1 p p
λ h(u)u dσ + |∇u| + |u| dx − p Ω r ∂Ω ∂Ω   1 1 − |∇u|p + |u|p dx. = p r Ω 



|u|r dσ ∂Ω r r |u| dσ

H (u) dσ −

Nonlinear boundary conditions

337

So, as r > p, we obtain uW 1,p (Ω)  C = C(c0 , p, r). Now by the Sobolev trace inequality, we get uLs (∂Ω)  S −1/p uW 1,p (Ω)  C = C(c0 , p, r, s, Ω).

(3.22)

Let us define uL (x) =




u(x)  L, u(x) > L.

u(x), L, pβ

Multiplying (3.20) by uL u we get 

Ω

pβ  |∇u|p−2 ∇u∇ uL u dx +

=λ



pβ

∂Ω

h(u)uuL dσ +





pβ

Ω

up uL dx pβ

∂Ω

ur uL dσ.

Therefore, using that h(u)u  K q−r ur and the definition of uL , we obtain 

Ω

pβ |∇u|p uL

dx +



Ω

 dx  λK q−r + 1

pβ up uL



pβ

∂Ω

ur uL dσ.

β

Now we set wL = uuL . Then, we obtain p

wL W 1,p (Ω)  = |∇wL |p + |wL |p dx Ω

C C



Ω

pβ |∇u|p uL

Ω

pβ |∇u|p uL





 C λK q−r + 1

dx +



dx +

∂Ω



Ω



Ω

p(β−1) up β p uL |∇uL |p dx pβ up uL

dx



+



Ω

pβ up uL

dx



pβ

ur uL dσ.

Therefore, by Hölder and Sobolev trace inequalities, we get

 p p wL Lp∗ (∂Ω)  S −1 wL W 1,p (Ω)  C λK q−r + 1

 C λK q−r + 1

 

∂Ω

p∗

u

dσ



∂Ω

pβ

ur uL dσ

(r−p)/p∗ 

∂Ω

∗ wLα

dσ

p/α ∗

,

338

J.D. Rossi

where α ∗ =

pp∗ p∗ −r+p

< p∗ . So by (3.22),

 r−p p p wL Lp∗ (∂Ω)  C λK q−r + 1 uLp∗ (∂Ω) wL Lα∗ (∂Ω)

 p  C λK q−r + 1 wL Lα∗ (∂Ω) . ∗

Now if uβ+1 ∈ Lα (∂Ω), by the dominated convergence theorem and Fatou’s lemma, we p p get uβ+1 Lp∗ (∂Ω)  C(λK q−r + 1)uβ+1 Lα∗ (∂Ω) , that is,

Let κ =

(β+1)/p uLp∗ (β+1) (∂Ω)  C λK q−r + 1 uLα∗ (β+1) (∂Ω) .

p∗ α∗ .

Iterating the last inequality we have

θ uLκ j α∗ (∂Ω)  C λK q−r + 1 uLα∗ (∂Ω) .

Using again (3.22) we get uL∞ (∂Ω)  C(λK q−r + 1)θ . Hence, if K0 > C, for every K
K0 , there exists λ(K) such that if λ < λ(K) then uL∞ (∂Ω)  K. This result finishes the proof.  Now, we give a nonexistence result for (3.1) in the half-space RN + = {x1 > 0} (see also ) that shows that existence may fail when one considSection 4 for more results in RN + ers critical or subcritical growth in an unbounded domain. This nonexistence result is a consequence of a Pohozaev-type identity. N N q 2 T HEOREM 3.7. Let f satisfy (3.3) with q  p∗ . Let u ∈ W 1,p (RN + ) ∩ C (R+ ) ∩ L (∂R+ ) be a nonnegative solution of (3.1) such that

  ∇u(x)|x|N/p → 0 as |x| → +∞.

Then u ≡ 0.

We remark that the decay hypothesis at infinity is necessary, because for p = 2 u(x) = e−x1 is a solution of (3.1) for every q. P ROOF OF T HEOREM 3.7. First, we multiply the equation by u and integrate by parts to obtain   p p |∇u| + u dx − uq dx ′ = 0. (3.23) RN +

∂RN +

Note that our decaying and integrability assumptions on u justify all the integrations by parts made along this proof.

Nonlinear boundary conditions

339

Now we multiply by x∇u and integrate by parts to obtain −



RN +

|∇u|p−2 ∇u∇(x∇u) dx +



∂RN +

uq−1 x∇u dx ′ =

1 p



RN +

x∇up dx.

Hence, further integrations by parts gives us     N N −1 N −1 + |∇u|p dx − uq dx ′ = up dx. N N N p q p R+ ∂R+ R+ Using (3.23) we arrive at     N −1 N 2N q ′ −1 + − u dx = −1 + up dx > 0. N N p q p ∂R+ R+ Therefore, if u is not identically zero, we must have q > p∗ = p(N − 1)/(N − p) as we wanted to show.   Now we study a convex–concave problem with a nonlinear boundary condition. We follow [81] and refer to that paper for the proofs, see also [107]. In [108] a similar problem is studied. We study the existence of nontrivial solutions for the following problem 

−
u + u = |u|p−2 u in Ω, ∂u ∂ν

= λ|u|q−2 u

(3.24)

on ∂Ω.

The study of existence when the nonlinear term is placed in the equation, that is if one considers a problem of the form −
u = f (u) with Dirichlet boundary conditions, has received considerable attention, see, for example, [22,80] etc. We want to remark that we are facing two nonlinear terms in problem (3.24), one in the equation |u|p−2 u, and one in the boundary condition |u|q−2 u. Our interest now is to analyze the interplay between both. By solutions to (3.24) we understand critical points of the associated energy functional (defined on H 1 (Ω))    1 λ 1 |∇u|2 + |u|2 dx − |u|p dx − |u|q dσ. F(u) = 2 Ω p Ω q ∂Ω This functional F is well defined and C 1 in H 1 (Ω) if p and q verify 1 < q  2∗ =

2(N − 1) N −2

and 1 < p  2∗ =

2N . N −2

We look for conditions that ensure the existence of nontrivial solutions of (3.24), focusing our attention on the existence of positive ones. We distinguish several cases.

340

J.D. Rossi

Convex–concave subcritical case. We suppose that 1 < q < 2 < p.

(3.25)

We want to remark that the new feature of these problems is that we are facing a convex– concave problem where the convex nonlinearity appears in the equation and the concave one at the boundary condition. Notice that if we look at the positive solutions of these problems as the stationary states of the corresponding evolution equation since the righthand side of the equation represents a positive reaction term, and the boundary condition means a positive flux at the boundary, then some absorption is required to reach a nontrivial equilibrium. In our equation, this is the linear term +u. First, we assume that the exponents involved are subcritical, that is, 1 < q < 2∗ =

2(N − 1) N −2

and 1 < p < 2∗ =

2N . N −2

(3.26)

We have the following theorems that can be proved using standard variational arguments together with the Sobolev trace immersion that provides the necessary compactness. T HEOREM 3.8. Let p and q satisfy (3.25) and (3.26). Then there exists λ0 > 0 such that if 0 < λ < λ0 then problem (3.24) has infinitely many nontrivial solutions. Now we concentrate on positive solutions for (3.24). T HEOREM 3.9. Let p and q satisfy (3.25) and (3.26). Then there exists Λ > 0 such that there exist at least two positive solutions of (3.24) for every λ < Λ, at least one positive solution for λ = Λ, and there is no positive solution of (3.24) for λ > Λ. Moreover, there exists a constant C such that uL∞ (Ω)  C for every positive solution. Critical case. Next we analyze the existence of solution when we have a critical exponent p = 2∗ . Here we use again the concentration–compactness method introduced in [96,97] and follow some ideas from [80]. In these kind of problems the concentration is a priori possible on the boundary. This difficulty leads us to use technical estimates that are implicit in [98]. For 2 < q < 2∗ (notice that in this case q means a convex term) we have the following theorem. T HEOREM 3.10. Let p = 2∗ with 2 < q < 2∗ , then problem (3.24) has at least a positive nontrivial solution for every λ > 0. And for 1 < q < 2, the next one. T HEOREM 3.11. If p = 2∗ with 1 < q < 2, then there exists Λ such that problem (3.24) has at least two positive solutions for λ < Λ, at least one positive solution for λ = Λ and no positive solution for λ > Λ.

Nonlinear boundary conditions

341

Further results. Moreover, to obtain existence of solutions we can apply the implicit function theorem near λ0 = 0, u0 = 1 to get existence of solutions for any p and q, but imposing a restriction on the domain. We have the following result. T HEOREM 3.12. Given 1 < p < ∞, let Ω be a domain such that (p − 1) ∈ / σNeu (−
+ I ). Then, for any q ∈ (1, ∞), there exists λ0 > 0 such that, for every λ ∈ (0, λ0 ), there exists a positive solution uλ ∈ C α of (3.24) with uλ → 1 in C α as λ → 0. Here σNeu (−
+ I ) stands for the spectrum of −
+ I with homogeneous Neumann boundary conditions. Finally let us state a result for the remaining case, q = 2. In this case we have a bifurcation problem from the first eigenvalue of a related problem. Let λ1 be the first eigenvalue of
−
u + u = 0 in Ω, ∂u on ∂Ω. ∂ν = λu Notice that λ1 is just the best constant in the Sobolev trace embedding H 1 (Ω) ֒→ L2 (∂Ω) in the sense that λ1 u2L2 (∂Ω)  u2H 1 (Ω) . Then we have the following theorem. T HEOREM 3.13. Let q = 2 with 2 < p < 2∗ . Then there exists a positive solution of (3.24) if and only if 0 < λ < λ1 . These ideas can also be applied to 

−
u + u = λ|u|q−2 u ∂u p−2 u ∂ν = |u|

in Ω, on ∂Ω.

(3.27)

For this problem we assume that 1 < q < 2 < p, that is, p stands for the convex term, and q for the concave one, and p is subcritical (notice that in this case this means p < 2(N − 1)/(N − 2)). The results presented here for (3.24) have analogous statements for (3.27).  In [52] the system 

−
p u = λ|u|q−2 u + |u|r−2 u

δ|∇u|p−2 ∂u ∂ν

+ a(x)|u|p−2 u = 0

in Ω, on ∂Ω

is considered. Here δ ∈ {0, 1}, the function a is strictly positive and the exponents verify 1 < q < p < r < pN/(N − p). The author find the existence of two values λ1 < λ2 such that two branches of nonnegative solutions exist for λ ∈ (0, λ2 ) with the energy of one of them changing sign at λ1 .  In [31] Chipot, Fila and Quittner studied the problem



u = aup ∂u q ∂ν = u

in Ω, on ∂Ω,

(3.28)

342

J.D. Rossi

when p, q > 1, a > 0. The authors consider the existence of positive solutions, multiplicity, its symmetry, bifurcations (depending on the parameter a) and stability properties. They give a very complete picture in the one-dimensional case and extend some of the results to the general case, see also [32] for a more detailed study of the multidimensional case. Let us present the results for the one-dimensional problem. As it can be noticed the existence and symmetry of the solutions depend strongly on p and q. T HEOREM 3.14. Denote by E the set of solutions and Es the symmetric solutions of (3.28). For N = 1 (Ω = (−l, l)) we have five cases: (1) p > 2q − 1. Then card(E) = 1, E = Es , for any a > 0. (2) p = 2q − 1. Then E = ∅ for 0 < a  q and card(E) = 1, E = Es , for a > q. (3) q < p < 2q − 1. Then there exist 0 < a0 < a1 such that: E = ∅ for a < a0 ; card(E) = 1, E = Es , for a = a0 ; card(E) = 2, E = Es , for a0 < a  a1 ; card(E)
4, even, card(Es ) = 2 for a > a1 . (4) p = q. Then there exists 0 < a1 such that: E = ∅ for 0 < a  1/ l; card(E) = 1, E = Es , for 1/ l < a  a1 ; card(E) = 3, card(Es ) = 1 for a > a1 . (5) p < q. Then there exists 0 < a1 such that: card(E) = 1, E = Es for 0 < a  a1 ; card(E) = 3, card(Es ) = 1 for a > a1 .  In [121] and [122] Umezu studied the problem


−
u + c(x)u = λf (u) in Ω, a(x) ∂u ∂ν + b(x)g(u) = 0 on ∂Ω.

It is discussed there the existence and uniqueness of a branch of positive solutions. The main tools used here are not variational, the results are obtained via the implicit function theorem and ideas relying on super- and subsolutions.  In [36,37,103] the problem, 



− div a(x)|∇u|p−2 ∇u + h(x)ur−1 = f (λ, x, u)

a(x)|∇u|p−2 ∂u ∂ν

+ b(x)up−1

= θg(x, u)

in Ω, on ∂Ω,

is studied using a bifurcation approach with parameters λ and θ . The authors consider the subcritical case, p < r < pN/(N − p). They also consider the case of an unbounded domain using appropriate weighted Sobolev spaces.

4. Problems in RN + In this section we consider solutions of elliptic problems with nonlinear boundary conditions in RN +. We have already find a nonexistence result in RN + for subcritical exponents, Theorem 3.7.

Nonlinear boundary conditions

Now, let us look for positive solutions of the problem  −
u = aup in RN +, = uq − x∂u N

on {xN = 0}.

343

(4.1)

Here a
0 and p, q > 1. It is easy to see that for N = 1 solutions do not exist, also for N = 2 there is no positive solution of (4.1). When N
3 solutions exist for the critical exponents p = (N + 2)/(N − 2) and q = N/(N − 2). It was proved in [33] and [93] using the method of moving spheres (instead of moving planes) that in this case any positive solution has the form u(x) =

α , (|x − x| ˜ 2 + β)(N −2)/2

where α > 0, x˜ = (x˜1 , . . . , x˜N ) ∈ RN with x˜N = −α 2/(N −2) /(N − 2) and β = aα 4/(N −2) / (N (N − 2)). When a = 0 positive solutions do not exist if q < N/(N − 2), see [87]. Here the moving planes technique is applied with a plane parallel to the xN direction to prove that the solution has to depend only in the xN variable and therefore is the zero solution. See also [117] for symmetry properties of problems with nonlinear boundary conditions in RN + obtained via the moving planes device. In [30] it is proved that positive solutions do exist when p
(N + 2)/(N − 2) and q
N/(N − 2). Also it is proved there that solutions do not exist in any of the following cases: (i) p  (N + 2)/(N − 2), q  N/(N − 2) with at least one strict inequality, (ii) p < N/(N − 2), (iii) q < N/(N − 1). Moreover, explicit solutions exist when p = q > N/(N − 2) and a > 0. They have the form u(x) =

α , (|x − x| ˜ 2 )2/(p−1)

where x˜N = − a1 (N − 2p/(p − 1)) and α = (−2x˜N /(p − 1))1/(p−1) . The proof of nonexistence in case (i) follows by an application of the moving planes method, in cases (ii) and (iii) it is a consequence of some blow-up results for parabolic problems. The idea of existence is as follows, consider the auxiliary nonlinear eigenvalue problem   −
u = λu + a|u|p−1 u in B + (0, R) = |x| < R ∩ RN  +,   ∂u q−1 − x = |u| u on |x| < R, xN = 0 ,    N u=0 on |x| = R, xN > 0 .

The proof comes from the use of Rabinowitz’s theorem that shows that there exists a branch of solutions (uR , λR ) emanating from (0, λ1 ), 0 < λ < λ1 , λ1 the first eigenvalue

344

J.D. Rossi

of the corresponding linear problem. The fact that λR is positive is proved by a Pohozaev identity and used to show that λR → 0 as R → ∞. It is the Pohozaev identity where p
(N + 2)/(N − 2) and q
N/(N − 2) are needed. Then, Schauder estimates show that uR converges along a subsequence to a solution of (4.1) as R → ∞. The proof of existence finishes by showing some symmetry and monotonicity results that allow us to conclude that the limit u is nontrivial.  Escobar [53] and Beckner [18] proved the following sharp trace inequality in the half-space for the critical exponent, 

∂RN +

|u|2(N −1)/(N −2) dσ

(N −2)/(N −1)

(1/(N −1)  ' Ŵ(N − 1) 1 |∇u|2 dx . √ π(N − 2) Ŵ((N − 2)/2) RN +

This result has applications in geometry to the Yamabe problem on manifolds with boundary, see Section 5.  In [106] Park proved the following logarithmic Sobolev trace inequality 

∂RN +

|u|2 ln |u| dσ 

  N |∇u|2 dx . ln AN 2 RN +

The result is obtained as a limit case for the best Sobolev trace inequality proved by Escobar in [53], see Section 5. Also bounds for the best constant in this logarithmic inequality are proved in [106].  See also Section 12 and [42] for existence and symmetry results for a problem with a non-Lipschitz nonlinearity in the half-space.  Finally, in [24] Cabre and Sola-Morales studied existence and uniqueness of layer solutions for the problem, 


u = 0

= f (u) − x∂u N

in RN +,

on {xN = 0}.

(4.2)

For layer solutions we mean solutions of the form u(x1 , . . . , xN ) that are monotone increasing from −1 to 1 in the xi variable with i = N . 5. Yamabe problem on manifolds with boundary In this section we describe some results obtained by Escobar [56] concerning the problem of prescribing the mean curvature of the boundary for a Riemannian manifold. See also [3]. This problem leads naturally to an elliptic problem with nonlinear boundary conditions. We will only present the geometrical motivation and some results, refereing to [3,5, 53–56,60,85,86] for further references.

Nonlinear boundary conditions

345

Let (M N , g) be a compact Riemannian manifold with boundary, N
3. Let g˜ = be a conformally related metric to g. Now we address the following problem: Given a function f on the boundary of M, can you find a conformally related metric g˜ such that the scalar curvature of g˜ is zero on M and the mean curvature of ∂M with respect to the metric g˜ is the function f ? This problem is equivalent to finding a smooth positive function u defined on M such that  N −2 on M,
u − 4(N −1) Ru = 0 (5.1) N −2 N −2 ∂u N/(N −2) on ∂M, ∂ν + 2 hu = 2 f u

u4/(N −2) g

where R is the scalar curvature of M, h is the mean curvature of ∂M, and ν is the outer normal vector with respect to the metric g. If we choose g˜ = u4/(N −2) g, the first equation in (5.1) says that the metric g˜ has zero scalar curvature and the second equation says that the boundary has mean curvature f with respect to the metric g. ˜  we define the energy associated to our problem as For a function v ∈ C 1 (M) E(v) =

   N −2 N −2 |∇v|2 + Rv 2 + hv 2 , 4(N − 1) 2 M ∂M

and the Sobolev quotient Q(M, ∂M) by

where



  , v ≡ 0 on ∂M , Q(M, ∂M) = inf Q(v): v ∈ C 1 M E(v) . Q(v) =  ( ∂M |v|2(N −1)/(N −2) )(N −2)/(N −1)

The constraint set C(M) is defined as




 1  2(N −1)/(N −2) C(M) = v ∈ C M : f |v| =1 . ∂M

We have the following proposition. P ROPOSITION 5.1. There is a function u that realizes the minimum energy in C(M) if (max f )(N −2)/(N −1) inf E(v) < v∈C

N − 2 N −1 1/(N −1) Vol S . 2

The constant that appear in the right-hand side is the Sobolev constant of the ball in the Euclidean space. Now we assume the generic condition that there exists a point x0 ∈ ∂M where the eigenvalues of the second fundamental form at x0 are not the same. That is to say that x0 is a nonumbilic point.

346

J.D. Rossi

T HEOREM 5.1. Let M be an N -dimensional compact Riemannian manifold with boundary, N > 5. If M has a nonumbilic point on ∂M then

 Q(M, ∂M) < Q B(0, 1), ∂B(0, 1) ,

where B(0, 1) is the ball in the Euclidean space. Let f verifies that is somewhere positive and achieves a global maximum at a nonumbilic point x0 where
f (x0 )  C(N)π − hg2 (x0 ) (π stands for the second fundamental form), then problem (5.1) has a solution. For manifolds with positive Sobolev quotient and ∂M umbilic we have the following theorem. T HEOREM 5.2. Let M be an N -dimensional compact Riemannian manifold with boundary, N
3 and Q(M, ∂M) > 0. Assume that M is locally conformally flat and ∂M umbilic. If M is not conformally diffeomorphic to the ball and f verifies that it is somewhere positive and achieves a global maximum at x0 ∈ ∂M with ∇ k f (x0 ) = 0 for k = 1, . . . , N − 2, then problem (5.1) has a positive smooth solution. Finally, we state a nonexistence result. T HEOREM 5.3. Let B be the N -dimensional Euclidean ball. Let X be the conformal vector field on ∂B. For a function f such that  ∇f X˙ = 0, ∂B

problem (5.1) has no positive solution. See also [85] and [86] for other existence results assuming that (B(0, 1), g) is of positive type.  In [5] it is shown that when the metric g is close to the standard metric in the ball then there exists a positive solution of (5.1).  In [60] and [61] some a priori estimates on the solutions are given.  See [124] for existence results for general elliptic operators, may be in nondivergence form.

6. Dependence of the best Sobolev trace constant on the domain In this section we deal with the dependence of the best Sobolev trace constant on the domain. First, we consider the family of domains given by contraction or expanding a fixed domain, that is, for µ > 0 we consider the family of domains Ωµ = µΩ = {µx; x ∈ Ω}.

Nonlinear boundary conditions

347

The main purpose of this section is to describe the asymptotic behavior of the best Sobolev trace constants S(Ωµ , p, q) as µ → 0+ and µ → +∞. As we have mentioned in the preliminaries, for any 1 < p < N and 1 < q  p∗ = p(N − 1)/(N − p), we have that W 1,p (Ω) ֒→ Lq (∂Ω) and hence the following inequality holds p

p

SuLq (∂Ω)  uW 1,p (Ω) for all u ∈ W 1,p (Ω). This is known as the Sobolev trace embedding theorem. The best constant for this embedding is the largest S such that the above inequality holds, that is, S(Ω, p, q) =

inf

1,p

u∈W 1,p (Ω)\W0 (Ω)



p Ω|∇u|

+ |u|p dx . ( ∂Ω |u|q dσ )p/q

Moreover, if 1 < q < p∗ the embedding is compact and as a consequence we have the existence of extremals, that is, functions where the infimum is attained, see [71]. These extremals are weak solutions of the following problem 


p u = |u|p−2 u |∇u|p−2 ∂u ∂ν

= λ|u|q−2 u

in Ω, on ∂Ω.

Standard regularity theory, like the one sketched in Section 2, see also [119], and the strong 1,α  (Ω) ∩ C α (Ω) maximum principle [123], show that any extremal u belongs to the class Cloc and that is strictly one signed in Ω, so we can assume that u > 0 in Ω. In [46] Flores and del Pino, performed a detailed analysis of the behavior of extremals and best Sobolev constants in expanding domains for p = 2 and q > 2. Their first result says that the best Sobolev trace constant in an expanding domain approaches the one in the half-space and gives an estimate of the error. T HEOREM 6.1. There exists a constant γ = γ (q, N ) > 0 such that the following expansion holds   1

N 1 S(µΩ, 2, q) = S R , 2, q − γ max H (x) + o µ x∈∂Ω µ as µ → +∞. Here H (x) denotes the mean curvature of the boundary at x.

The second result proved in that paper says that the extremals constitute a single bump at the boundary, whose shape is asymptotically that of an extremal for the half-space trace theorem. This bump is centered around a point of maximum mean curvature. T HEOREM 6.2. Let yµ be a maximum point of uµ . Then xµ = yµ /µ ∈ ∂Ω verifies H (xµ ) → max H (x) x∈∂Ω

348

J.D. Rossi

as µ → +∞. Also, there exist constants α, β > 0 such that uµ (y)  αe−β|y−yµ | for all y ∈ µΩ. Besides, given a sequence µn → ∞ there exists a subsequence, an extremal w of S(RN , 2, q) and a rotation Q such that 

 sup unk (y) − w Q(y − yµnk )  → 0

y∈µnk Ω

as k → ∞.

The main ingredient of the proof is to obtain some information on the limit problem 


w = w ∂w ∂ν

= wp

in RN +,

on ∂RN +.

For this problem we have the existence of a least energy solution that verifies the decay estimate |w(x)| + |∇w(x)|  C1 e−C2 |x| . See also Section 4 for more results for problems with nonlinear boundary conditions in RN +. Let us go back to our general problem W 1,p (Ωµ ) → Lq (∂Ωµ ). Now we deal with the case of contractions, µ → 0+. As we will see the behavior of the Sobolev constant and extremals is very different when the domain is contracted than when it is expanded. Let us call uµ an extremal corresponding to Ωµ . Making a change of variables, we go back to the original domain Ω. If we define vµ (x) = uµ (µx), we have that vµ ∈ W 1,p (Ω) and  −p p p (N q−Np+p)/q Ω µ  |∇vµ | + |vµ | dx . (6.1) S(Ωµ , p, q) = µ ( ∂Ω |vµ |q dσ )p/q We can assume, and we do so, that the functions uµ are normalized so that 

∂Ω

|vµ |q dσ = 1.

We remark that the quantity (6.1) is not homogeneous under dilations or contractions of the domain. This is a remarkable difference with the study of the Sobolev embedding 1,p W0 (Ω) ֒→ Lq (Ω). The first result of Fernández Bonder and the author in [74] is the following theorem. T HEOREM 6.3. Let 1 < q < p∗ . Then S(Ωµ , p, q) |Ω| , = (N q−Np+p)/q µ→0+ µ |∂Ω|p/q lim

(6.2)

and if we scale the extremals uµ to the original domain Ω as vµ (x) = uµ (µx), x ∈ Ω, with vµ Lq (∂Ω) = 1, then vµ is nearly constant in the sense that vµ → |∂Ω|−1/q in W 1,p (Ω).

Nonlinear boundary conditions

349

Observe that the behavior of the Sobolev trace constant, strongly depends on p and q. If we call βpq = (N q − Np + p)/q then we have that, as µ → 0+, 0 if βpq > 0,  if βpq < 0, S → +∞  C = 0 if βpq = 0.

Let us remark that the influence of the geometry of the domain appears in (6.2). An idea of the proof of Theorem 6.3 runs as follows. Let us begin by the simple observation that, taking u ≡ 1 as a test function in (6.1), it follows that S(Ωµ , p, q)  µ(N q−Np+p)/q

|Ω| . |∂Ω|p/q

(6.3)

This shows that the ratio S(Ωµ , p, q)/µ(N q−Np+p)/q is bounded. So a natural question will be to determine if it converges to some value. This is answered in Theorem 6.3 that we prove next. P ROOF OF T HEOREM 6.3. Let uµ ∈ W 1,p (Ωµ ) be a extremal for S(Ωµ , p, q) and define vµ (x) = uµ (µx), we have that vµ ∈ W 1,p (Ω). We can assume that the functions uµ are chosen so that 

∂Ω

|vµ |q dσ = 1.

Equations (6.1) and (6.3) give, for µ < 1, p vµ W 1,p (Ω)





Ω

µ−p |∇vµ |p + |vµ |p dx 

|Ω| , |∂Ω|p/q

so there exists a function v ∈ W 1,p (Ω) and a sequence µj → 0+ such that vµj ⇀ v

weakly in W 1,p (Ω),

vµj → v

in Lp (Ω),

vµj → v

in Lq (∂Ω).

Moreover, 

Ω

|∇vµ |p dx 

|Ω| µp . |∂Ω|p/q

350

J.D. Rossi

Hence ∇vµ → 0 in Lp (Ω). It follows that the limit v is a constant and must verify  q −1/q and so the full sequence v converges weakly µ ∂Ω |v| = 1, hence v = const = |∂Ω| in W 1,p (Ω) to v. From our previous bounds we have vµ →

1 |∂Ω|1/q

in Lp (Ω)

and



Ω

|∇vµ |p dx → 0.

Therefore, we have strong convergence, vµ → |∂Ω|−1/q in W 1,p (Ω). The proof is finished.  In the special case p = q, the problem, 


p u = |u|p−2 u |∇u|p−2 ∂u ∂ν

in Ω,

= λ|u|q−2 u

on ∂Ω,

(6.4)

becomes a nonlinear eigenvalue problem. For p = 2, this eigenvalue problem is known as the Steklov problem [115]. In [71] it is proved, applying the Ljusternik–Schnirelman critical point theory on C 1 manifolds, that there exists a sequence of variational eigenvalues λk ր +∞, see Section 3. It is easy to see that the first eigenvalue λ1 (Ω) verifies λ1 (Ω) = S(Ω, p, p). So Theorem 6.3 shows a difference in the behavior of the first eigenvalue of (6.4) with p = q with respect to the domain with the behavior of the first eigenvalue of the following Dirichlet problem


−
p u = λ|u|p−2 u in Ω, u=0 on ∂Ω,

where it is a well-known fact that λ1 increases as the domain decreases, see [79]. Recall from Section 3 that variational eigenvalues λk of (6.4) are characterized by p

uLp (∂Ω) 1 , = sup min λk C∈Ck u∈C up 1,p W

(Ω)

where Ck = {C ⊂ W 1,p (Ω); C is compact, symmetric and γ (C)
k} and γ is the genus. It is shown in [72] that there exists a second eigenvalue for (6.4) and that it coincides with the second variational eigenvalue λ2 . Moreover, the following characterization of the second eigenvalue λ2 holds λ2 = inf




u∈A

Ω

|∇u| + |u| dx , p

p

where A = {u ∈ W 1,p (Ω); uLp (∂Ω) = 1 and |∂Ω ± |
c}, with ∂Ω + = {x ∈ ∂Ω; u(x) > 0} and ∂Ω − is defined analogously. Concerning the eigenvalue problem, we have the following result.

Nonlinear boundary conditions

351

T HEOREM 6.4. There exists a constant λ˜ 2 such that lim µp−1 λ2 (Ωµ ) = λ˜ 2 .

µ→0+

This constant λ˜ 2 is the first nonzero eigenvalue of the following problem 


p u = 0

|∇u|p−2 ∂u ∂ν

in Ω, ˜ p−2 u = λ|u|

on ∂Ω.

(6.5)

Moreover, if we take an eigenfunction u2,µ associated to λ2 (Ωµ ) and scale it to Ω as in Theorem 6.3, we obtain that v2,µ → v˜2 in W 1,p (Ω), where v˜2 is an eigenfunction of (6.5) associated to λ˜ 2 . Also, every eigenvalue λ2 (Ωµ )  λ(Ωµ )  λk (Ωµ ) of (6.4) (variational or not) behaves as λ(Ωµ ) ∼ µ1−p as µ → 0+. Finally, if µj → 0 and λj = λ(Ωµj ) is a sequence of eigenvalues such that there exists λ with p−1

lim µj

j →∞

λj = λ.

Let (vj ) be the sequence of associated eigenfunctions rescaled as in Theorem 6.3, then (vj ) has a convergent subsequence (vjk ) and a limit v, that is, an eigenfunction of (6.5) with eigenvalue λ. Observe that the first eigenvalue of (6.5) is zero with associated eigenfunction a constant. Hence, Theorem 6.3 says that the first eigenvalue and the first eigenfunction of our problem (6.4) converges to the ones of (6.5). Theorem 6.4 says that λ(Ωµ ) → +∞ as µ → 0+ for the remaining eigenvalues and that problem (6.5) is a limit problem for (6.4) when µ → 0+. Since we are dealing with the eigenvalue problem, let us prove the isolation and simplicity for the first eigenvalue of the p-Laplacian with a nonlinear boundary condition. This result was proved by Martinez and the author and is contained in [100]. So, let us study the first eigenvalue for the following problem 


p u = |u|p−2 u

|∇u|p−2 ∂u ∂ν

= λ|u|p−2 u

in Ω, on ∂Ω.

(6.6)

We have the following result, similar to the one known for Dirichlet boundary conditions [7]. T HEOREM 6.5. λ1 is isolated and simple. We remark that this theorem says that the extremals of the Sobolev trace inequality are unique up to multiplication by a real number. In the special case of a ball, Ω = B(0, R), our result implies that the first eigenfunction is radial. In fact, if u1 (x) is an eigenfunction associated to λ1 and θ (x) is any rotation then u1 (θ (x)) is also an eigenfunction, by our

352

J.D. Rossi

result we have that u1 (x) = u1 (θ (x)). We conclude that u1 must be radial. Also from our results it follows that any other eigenvalue has nonradial eigenfunctions as they have to change sign on the boundary (see Lemma 6.4). Now we prove Theorem 6.5. To clarify the exposition, we will divide the proof in several lemmas. L EMMA 6.1. Let u1 be an eigenfunction with eigenvalue λ1 , then u1 does not change sign  on Ω. Moreover, if u1 is C 1,α (Ω), it does not vanish on Ω. P ROOF. Recall that λ1 =

inf




u∈W 1,p (Ω)

Ω

|∇u|p + |u|p dx,



∂Ω

|u|p dσ = 1 .

(6.7)

Hence if u1 is a minimizer, we have that |u1 | is also a minimizer of (6.7). By the maximum principle (see [123]), we have that |u1 | > 0 in Ω. Assume that u1 is regular and that there exists x0 ∈ ∂Ω such that u1 (x0 ) = 0. By the Hopf lemma (see [123]) we have that 1| the normal derivative has strict sign, ∂|u ∂ν (x0 ) < 0, but the boundary condition imposes ∂|u1 |   ∂ν (x0 ) = 0, a contradiction, that proves that |u1 | > 0 in Ω. The result follows. Now we state an auxiliary lemma, whose proof can be found in [95]. L EMMA 6.2. (a) Let p
2. Then for all ξ1 , ξ2 ∈ RN , |ξ2 |p
|ξ1 |p + p|ξ1 |p−2 ξ1 , ξ2 − ξ1  + C(p)|ξ1 − ξ2 |p . (b) Let p < 2. Then for all ξ1 , ξ2 ∈ RN , |ξ2 |p
|ξ1 |p + p|ξ1 |p−2 ξ1 , ξ2 − ξ1  + C(p)

|ξ1 − ξ2 |p , (|ξ2 | + |ξ1 |)2−p

where C(p) is a constant depending only on p. L EMMA 6.3. λ1 is simple. Let u, v be two eigenfunctions associated with λ1 , then there exists c such that u = cv. P ROOF. By Lemma 6.1 we can assume that u, v are positive in Ω. We perform the fol to obtain our result we lowing calculations assuming that u, v are strictly positive in Ω, can consider u + ε and v + ε and let ε → 0 at the end as in [95]. Therefore we can take η1 = (up − v p )/up−1 and η2 = (v p − up )/v p−1 as test functions in the weak form of (6.6) satisfied by u and v, respectively. We have  p  u − vp p−2 dx |∇u| ∇u∇ up−1 Ω  p  p   u − vp u − vp p−2 p−2 =λ dσ − dx |u| u |u| u up−1 up−1 ∂Ω Ω

Nonlinear boundary conditions

353

and v p − up dx |∇v| ∇v∇ v p−1 Ω  p  p   v − up v − up p−2 p−2 =λ dσ − dx. |v| v |v| v v p−1 v p−1 ∂Ω Ω





p−2

Adding both equations we get

0=



Ω

|∇u|p−2 ∇u∇



up − v p up−1



dx +



Ω

|∇v|p−2 ∇v∇



v p − up v p−1



dx. (6.8)

Using that 

up − v p ∇ up−1



= ∇u − p

v p−1 vp ∇v + (p − 1) ∇u, up up−1

we obtain that the first term of (6.8) is  p  u − vp p−2 dx |∇u| ∇u∇ up−1 Ω    v p−1 vp p p−2 = |∇u|p dx |∇u| − p |∇u| ∇v∇u dx + (p − 1) p−1 up Ω Ω u Ω   = |∇ ln u|p up dx − p v p |∇ ln u|p−2 ∇ ln u, ∇ ln vuv dx Ω

+



Ω

Ω

(p − 1)|∇ ln u|p v p dx.

We also have an analogous expression for the second term of (6.8). Using both expressions we get that (6.8) becomes 

p   u − v p |∇ ln u|p − |∇ ln v|p dx 0= Ω

−p −p



Ω

v p |∇ ln u|p−2 ∇ ln u, ∇ ln v − ∇ ln u dx

Ω

up |∇ ln v|p−2 ∇ ln v, ∇ ln u − ∇ ln v dx.



Taking ξ1 = ∇ ln u and ξ2 = ∇ ln v and using Lemma 6.2 we get, for p
2, 

 C(p)|∇ ln u − ∇ ln v|p up + v p dx. 0
Ω

354

J.D. Rossi

Hence, 0 = |∇ ln u − ∇ ln v|. This implies that u = kv as we wanted to prove. For p < 2 we use the second part of Lemma 6.2 as above.  Now we turn our attention to the proof of the isolation of the first eigenvalue. In order to prove this we need the following nodal result. L EMMA 6.4. Let w be an eigenfunction corresponding to λ = λ1 . Then w changes sign on ∂Ω, that is, w + |∂Ω = 0 and w − |∂Ω = 0. Moreover, there exists a constant C such that  − ∂Ω 
Cλ−β ,

 + ∂Ω 
Cλ−β ,

(6.9)

where ∂Ω + = ∂Ω ∩ {w > 0}, ∂Ω − = ∂Ω ∩ {w < 0}, β = (N − 1)/(p − 1) if 1 < p < N and β = 2 if p
N . Here |A| denotes the (N − 1)-dimensional measure of a subset A of the boundary. P ROOF. Assume that w does not change sign in Ω, then we can assume that w > 0 in Ω using ideas similar to those of Lemma 6.1. Let u1 be a positive eigenfunction associated to λ1 . Making similar computations as the ones performed in the proof of Lemma 6.3 we arrive at   

p 

p p u1 − w dσ
C |∇ ln w − ∇ ln u1 |p u1 + w p dx
0. (λ1 − λ) Ω

∂Ω

Therefore if we take kw instead of w we get that, for every k > 0, we have 

∂Ω



p u1 − k p w p dσ  0,

a contradiction if we take   p p k w dσ < ∂Ω

∂Ω

p u1 dσ

.

Therefore w changes sign in Ω and by the maximum principle [123], also w changes sign in ∂Ω. Let us use w − as test function in the weak form of (6.6) satisfied by w to obtain 

Ω

|∇w − |p dx +



|w − |p dx = λ

Ω



∂Ω∩{w λ1 such that λ1 is the unique eigenvalue in [0, a]. P ROOF. From the characterization of λ1 it is easy to see that λ1  λ for every eigenvalue λ. Assume that λ1 is not isolated, then there exists a sequence λk with λk > λ1 , λk ց λ1 . Let wk be an eigenfunction associated to λk , we can assume that wk W 1,p (Ω) = 1. Therefore we can extract a subsequence (that we still denote by wk ) such that wk → u1 in Lp (∂Ω). Let us define φk ∈ (W 1,p (Ω))′ as  φk (u) = λk |wk |p−2 wk u dσ ∂Ω

and φ ∈ (W 1,p (Ω))′ by  |u1 |p−2 u1 u dσ. φ(u) = λ1 ∂Ω

From the Lp (∂Ω) convergence of wk to u1 we get that φk converges to φ in (W 1,p (Ω))′ . Using the continuity of Ap given by Lemma 3.1 we get that the sequence wk converge strongly in W 1,p (Ω). Therefore, passing to the limit in the weak form of (6.6) we get that u1 is an eigenfunction with eigenvalue λ1 . By Lemma 6.1 we can assume that u1 > 0 on ∂Ω. By Egorov’s theorem we can find a subset Aε of ∂Ω such that |Aε | < ε and wk → u1 > 0 uniformly in ∂Ω \ Aε . This contradicts the fact that, by (6.9), we have, −(N −1)/(p−1) for every k, |∂Ωk− | = ∂Ω ∩ {wk < 0}
Cλk . This result completes the proof.  Now, we go back to our original problem, the asymptotic behavior of the best Sobolev trace constant when considered over the family Ωµ . Let us look now to the case µ → +∞. In this case we find, as before, that the behavior strongly depends on p and q. We prove the following theorem. T HEOREM 6.6. Let βpq = (qN − pN + p)/q. It holds: (1) If 1 < q < p, 0 < c1 µβpq −1  S(Ωµ , p, q)  c2 µβpq −1 . (2) If p  q < p∗ , 0 < c1  S(Ωµ , p, q)  c2 < ∞. For the lower bound in (2) in the case p < q < p∗ , we have to assume that the corresponding extremals vµ rescaled such that maxΩ vµ = 1 verify |∇vµ |  Cµ. Moreover, for all cases, we have that the corresponding extremals uµ rescaled as in Theorem 6.3 concentrates at the boundary, in the sense that  |vµ |p dx  Cµ−βpq → 0 as µ → +∞ if q
p, Ω

356

J.D. Rossi

and 

Ω

|vµ |p dx  Cµ−1 → 0

as µ → +∞ if q < p,

with 

∂Ω

|vµ |q dσ = 1.

As before, the behavior of the Sobolev trace constant depends on p and q. We have that, as µ → +∞, S→0

if βpq − 1 < 0, i.e., q < p,

0 < c1  S  c 2 < ∞

if βpq − 1
0, i.e., q
p.

The hypothesis |∇vµ |  Cµ is a regularity assumption, see [29] and [119] for regularity results. As a consequence we have that the extremals do not develop a peak if 1 < q < p as in this case we have that  c1  |vµ |p dσ  c2 ∂Ω

and 

∂Ω

|vµ |q dσ = 1.

For p = q it is proved in [100] that the first eigenvalue λ1 (Ωµ ) = S(Ωµ , p, p) is isolated and simple, see Theorem 6.5. As a consequence of this if Ω is a ball, the extremal vµ is radial and hence it does not develop a peak. Finally, for q > p the extremals develop peaking concentration phenomena in the sense that, for every a > 0, a p |∂Ω ∩ {vµ > a}| → 0 as µ → +∞, with maxΩ vµ = 1. This is in concordance with the results of [46] where for p = 2, q > 2 they find that the extremals concentrate, with the formation of a peak near a point of the boundary where the curvature maximizes. We believe that for q > p, extremals develop a single peak as in the case p = 2. Nevertheless that kind of analysis needs some fine knowledge of the limit problem in RN + that is not yet available for the p-Laplacian. Let us give an idea of the proof of the lower bounds. In the case p = q we can obtain the lower bound by an approximation procedure. We replace W 1,p (Ω) by an increasing sequence of subspaces in the minimization problem. Then we prove a convergence result and find a uniform bound from below for the approximating problems. We believe that this idea can be used in other contexts. For the case q > p we use our assumption |∇vµ |  Cµ to prove a reverse Hölder inequality for the extremals on the boundary that allows us to reduce to the case p = q. Finally, for large µ, in the case p = q we can prove that every eigenvalue is bounded.

Nonlinear boundary conditions

357

T HEOREM 6.7. Let λ1 (Ωµ )  λ(Ωµ )  λk (Ωµ ) be an eigenvalue of (6.4) in Ωµ (variational or not). Then there exists two constants, C1 , C2 > 0, independent of µ such that 0 < C1  λ(Ωµ )  C2 < +∞ for every µ large. Now we continue our study of the dependence of the best constant S(Ω, p, q) and extremals on the domain by considering the best Sobolev constant in thin domains. Now we consider a different family of domains. Let N = n + k and define the family  Ωµ = (µx, y) | (x, y) ∈ Ω, x ∈ Rn , y ∈ Rk .

Remark that for small values of µ, Ωµ is a narrow domain in the x direction. Our first result shows that, when the domain is very narrow, the problem of looking at the trace of a function is equivalent, in some sense, to the problem of the immersion of the function in the projection of the domain over the y variables. More precisely, we define the projection   P (Ω) = y ∈ Rk  ∃x ∈ Rn with (x, y) ∈ Ω

and consider the weighted Sobolev embedding W 1,p (P (Ω), α) ֒→ Lq (P (Ω), β) with associated best constant given by  p p

 P (Ω) (|∇v| + |v| )α(y) dy   inf Sα,β P (Ω), p, q = . ( P (Ω) |v|q β(y) dy)p/q v∈W 1,p (P (Ω),α)

We have the following theorem.

T HEOREM 6.8. Let 1  q < p∗ . Then there exist two nonnegative weights α, β ∈ L∞ (P (Ω)) such that lim

µ→0+

 S(Ωµ , p, q) = Sα,β P (Ω), p, q µ(nq−np+p)/q

and if we scale the extremals uµ of S(Ωµ , p, q) to the original domain Ω as vµ (x, y) = q uµ (µx, y), (x, y) ∈ Ω, normalized as uµ Lq (∂Ωµ ) = µn−1 , then vµ → v = v(y) strongly Sα,β (P (Ω), p, q). in W 1,p (Ω), where v ∈ W 1,p (P (Ω), α) is an extremal for  We want to remark that the weights α and β can be determined in terms of the geometry of Ω. In fact, α(y) = |Ωy | where Ωy is the section at level y of Ω. To clarify the content of the result, assume that Ω is a product, Ω = Ω1 × Ω2 where Ω1 ⊂ Rn and Ω2 ⊂ Rk . Then Ωµ = µΩ1 × Ω2 = {(µx, y) | x ∈ Ω1 , y ∈ Ω2 }. As in Theorem 6.8, let us call uµ an extremal corresponding to Ωµ and define vµ (x, y) = uµ (µx, y). We have that vµ ∈ W 1,p (Ω) and  −1 )|p + |vµ |p dx dy S(Ωµ , p, q) Ω |(µ ∇x vµ , ∇y vµ   , = ( ∂Ω1 ×Ω2 |vµ |q dσx dy + µ Ω1 ×∂Ω2 |vµ |q dx dσy )p/q µ(nq−np+p)/q

358

J.D. Rossi

where ∇x u = (ux1 , . . . , uxn ) and ∇y u = (uy1 , . . . , uyk ). The normalization imposed in Theorem 6.8 in this case reduces to   |vµ |q dx dσy = 1. (6.10) |vµ |q dσx dy + µ Ω1 ×∂Ω2

∂Ω1 ×Ω2

In this simpler case, the weight functions α, β are constants and can be computed explicitly, in fact, α(y) = |Ω1 | and β(y) = |∂Ω1 |. Hence, Theorem 6.8 reads as follows. T HEOREM 6.9. Let 1  q < p∗ and Ω = Ω1 × Ω2 . Then lim

µ→0+

S(Ωµ , p, q) |Ω1 |  = S(Ω2 ), |∂Ω1 |p/q µ(nq−np+p)/q

where  S(Ω2 ) =  S1,1 (Ω2 , p, q) is the usual Sobolev constant. Moreover, if we scale the extremals uµ to the original domain Ω as vµ (x, y) = uµ (µx, y), x ∈ Ω1 , y ∈ Ω2 , normalized by (6.10), then vµ → v = v(y) strongly in W 1,p (Ω), where v ∈ W 1,p (Ω2 ) is an extremal for  S(Ω2 ). Observe, that the critical exponent for the Sobolev embedding, W 1,p (Ω2 ) ֒→ Lq (Ω2 ) valid for 1  q < pk/(k − p), is larger than the one for the Sobolev trace embedding W 1,p (Ω) ֒→ Lq (∂Ω), which holds for 1  q < p(k + n − 1)/(k + n − p). Again, in the special case p = q, the problem becomes a nonlinear eigenvalue problem. Following [71] (see also [49]), a sequence of variational eigenvalues λj can be characterized by p

λj = inf max C∈Cj u∈C

uW 1,p (Ω) p

uLp (∂Ω)

,

(6.11)

where    Cj = Φ S j −1 ⊂ W 1,p (Ω)  Φ : S j −1 → W 1,p (Ω) \ {0} is continuous and odd

and S j −1 is the unit sphere of Rj . These eigenvalues differ slightly from the ones considered in [71]. However, the same arguments used there apply proving that in fact {λj } is an unbounded sequence of eigenvalues. When µ goes to zero, there is a limit problem which is a weighted eigenvalue problem on the projection P (Ω). Let α and β be the weights given by Theorem 6.8 and consider the following eigenvalue problem 

 p−2 v ¯ − div α|∇v|p−2 ∇v + α|v|p−2 v = λβ|v| ∂v ∂ν

=0

in P (Ω), on ∂P (Ω).

(6.12)

Nonlinear boundary conditions

359

For problem (6.12), one can define the sequence λ¯ j = inf max C∈Cj u∈C



where

p P (Ω) (|∇u|



+ |u|p )α dy

p P (Ω) |u| β dy

,

(6.13)

 



 Cj = Φ S j −1 ⊂ W 1,p P (Ω) : Φ : S j −1 → W 1,p P (Ω) \ {0} is continuous and odd .

Once again, applying the Ljusternik–Schnirelmann critical point theory one could check that {λ¯ j } is an unbounded sequence of eigenvalues for (6.12). However, this fact is a direct consequence of our next result. T HEOREM 6.10. Let λj,µ given by (6.11) in Ωµ and let uj,µ be an associated eigenfunction normalized as in Theorem 6.8. Then lim

µ→0

λj,µ = λ¯ j , µ

where λ¯ j is defined by (6.13) and is an eigenvalue of (6.12). Also, along a subsequence, vj,µ (x, y) = uj,µ (µx, y) converges strongly in W 1,p (Ω) to a function v¯j = v¯j (y) which is an eigenfunction of (6.12) with eigenvalue λ¯ j . Observe that the first eigenvalue λ1 coincides with the best Sobolev trace constant S(Ω, p, p). Hence, for p = q and for the first eigenvalue, Theorem 6.8 and Theorem 6.10 coincide. As before, in the case Ω = Ω1 × Ω2 , the limit problem has a simpler form, that is, 

−
p v + |v|p−2 v = ∂v ∂ν

|∂Ω1 | ¯ p−2 v |Ω1 | λ|v|

in Ω2 ,

=0

on ∂Ω2 .

However, Theorem 6.10 conserves the same statement. Our last result is concerned with the following fact: once the domain has been contracted in the x direction, we can now try to contract it in the y direction and see if the limit coincides with the one obtained by contracting the domain in every direction at the same time. Surprisingly, this is not the case. In fact, we obtain the following theorem. T HEOREM 6.11. Let Ω = Ω1 × Ω2 and consider Ωµ,ν = {(µx, νy): (x, y) ∈ Ω}, then lim



ν→0

S(Ωµ,ν , p, q) µ→0 µ(nq−np+p)/q ν (kq−kp)/q lim



=

|Ω| . (|∂Ω1 ||Ω2 |)p/q

360

J.D. Rossi

By the previous results, Theorem 6.3, we have lim

µ→0

S(µΩ, p, q) |Ω| |Ω| = . = |∂Ω|p/q (|∂Ω1 ||Ω2 |)p/q µ(N q−Np+p)/q

This shows that the double limit lim(µ,ν)→(0,0) S(Ωµ,ν , p, q) does not exists. For a general domain Ω we have lim

ν→0



S(Ωµ,ν , p, q) (nq−np+p)/q µ→0 µ ν (kq−kp)/q lim



|Ω| =  . ( P (Ω) β dy)p/q

To prove this fact we assume that the immersion W 1,p (P (Ω), α) ֒→ Lq (P (Ω), β) is compact. To see in which cases this holds, see [66].

7. Symmetry of extremals The aim of this section is to study of the following problem: Given a ball of radius ρ, B(0, ρ), in RN , N
3, decide whether or not there exists a radial extremal for the embedding



 H 1 B(0, ρ) ֒→ Lq ∂B(0, ρ) .

First, let us introduce our motivation. Recall that the best constant for the Sobolev trace embedding S(Ω, 2, q) =

inf

v∈H 1 (Ω)\H01 (Ω)



2 Ω |∇v|

+ |v|2 dx . ( ∂Ω |v|q dσ )2/q

(7.1)

As we noticed before, the best constant S(Ω, 2, q) is not homogeneous under dilatations. In fact, we have S(µΩ, 2, q) = µ

β

inf

v∈H 1 (Ω)\H01 (Ω)



Ω

µ−2 |∇v|2 + |v|2 dx  , ( ∂Ω |v|q dσ )2/q

where β = (N q − 2N + 2)/q. For 1  q < 2∗ = 2(N − 1)/(N − 2), the embedding is compact, so we have existence of extremals, that is, functions where the infimum is attained. These extremals are weak solutions of the following problem 


u = u ∂u q−2 u ∂η = λ|u|

in Ω, on ∂Ω.

The asymptotic behavior of S(µΩ, p, q) in expanding (µ → ∞) and contracting domains (µ → 0) was studied in the previous section, see also [46] and [73]. In [46] it is proved that

Nonlinear boundary conditions

361

for expanding domains and q > 2, S(µΩ, 2, q) → S(RN + , 2, q). In [73], see Theorem 6.3, it is shown that lim

µ→0+

S(µΩ, 2, q) |Ω| . = µβ |∂Ω|2/q

As we mentioned in the previous section, the behavior of the extremals for (7.1) in expanding and contracting domains is also studied in [46] and [73]. For expanding domains, it is proved in [46] that the extremals develop a peak near a point where the mean curvature of the boundary is a maximum. For contracting domains, we have that the extremals, when rescaled to the original domain as v(x) = u(µx), x ∈ Ω, and normalized with vLq (∂Ω) = 1, are nearly constant in the sense that limµ→0 v = |∂Ω|−1/q in H 1 (Ω). A big difference between the Sobolev trace theorem and the Sobolev embedding theorem arises in the behavior of extremals. Namely, if Ω is a ball, Ω = B(0, ρ), as the extremals do not change sign, from results of [82] the extremals for the usual Sobolev embedding, H01 (B(0, 1)) ֒→ Lq (B(0, 1)), are radial while, if q exceeds 2 and ρ is large, extremals for (7.1) are not, since they develop a peaking concentration phenomena as is described in [46]. The above discussion leads naturally to the purpose of this section: the study of the symmetry properties for the extremals of the Sobolev trace embedding in small balls. We find that the symmetry properties of the extremals depend on the size of the ball. Our first result describes when there exists a radial extremal. T HEOREM 7.1. Let 2∗ = 2(N − 1)/(N − 2) be the critical exponent for the Sobolev trace immersion. Concerning symmetry properties of the extremals for the embedding H 1 (B(0, ρ)) ֒→ Lq (∂B(0, ρ)) there holds: (1) Let 1 < q  2. For every ρ > 0 there exists a radial extremal. (2) Let 2 < q < 2∗ . There exists ρ0 > 0 such that, for every ρ < ρ0 , there is a unique positive extremal u, normalized such that u(ρx)Lq (∂B(0,1)) = 1; moreover, this extremal is a radial function. However, for large values of ρ, there is no radial extremal. (3) Let q = 2∗ . There exists ρ0 > 0 such that, for every ρ < ρ0 , there is a positive radial extremal. The main ingredient of the proof of the symmetry result for small balls is the implicit function theorem. We remark that the moving planes technique cannot be applied to obtain symmetry results in this case, as the extremals for large ρ are not radial. For small balls, using the symmetry result in balls, for the critical exponent 2∗ = 2 × (N − 1)/(N − 2), we can prove existence of extremals, which turns out to be radial functions. For general domains Ω, see [3], where it is proved that a extremal exists if the domain (bounded or not) verifies that it contains a point at the boundary with strictly positive curvature. We remark that the existence of extremals for the critical exponent is not trivial, this is due to the lack of compactness. This result has to be compared with the case ∗ of the immersion H01 (B(0, ρ)) → L2 (B(0, ρ)) where it is well known that, by Pohozaev identity, there is no positive solution regardless the size of the ball for the critical exponent 2∗ = 2N/(N − 2). However, there exist solutions for topologically nontrivial domains.

362

J.D. Rossi

In [92] Lami-Dozo and Torne studied the symmetry and symmetry breaking of the extremals in a ball for general 1 < p < ∞. Consider  p p

 Ω|∇u| + |u| dx S B(0, ρ), p, q = inf . (7.2) q p/q 1,p u∈W 1,p (Ω)\W (Ω) ( ∂Ω |u| dσ ) 0

As we have already mentioned for subcritical q, 1 < q < p(N − 1)/(N − p), the best constant is attained and we can  assume that extremals are positive in B(0, ρ). With the normalization (different from ∂Ω uq = 1)

 S B(0, ρ), p, q



uq dσ

∂B(0,ρ)

the extremals are solutions of 
p u = up−1 |∇u|p−2 ∂u ∂ν

(p−q)/q

= 1,

in B(0, ρ),

= λuq−1

on ∂B(0, ρ).

(7.3)

Let us call ϕ1 the eigenfunction associated with the first eigenvalue λ1 of the problem 


p ϕ = |ϕ|p−2 ϕ |∇ϕ|p−2 ∂ϕ ∂ν

in B(0, ρ),

= λ|ϕ|p−2 ϕ

on ∂B(0, ρ).

(7.4)

Recall that the first eigenvalue is isolated and simple, see Section 6 and [100], therefore ϕ1 is radial. We have the following theorem. T HEOREM 7.2. Let ρ > 0 and 1 < p < ∞ be fixed. (1) If there exists a radial minimizer of (7.2) then it is a multiple of ϕ1 . (2) Assume that there exists a radial minimizer for S(B(0, ρ), p, q0 ) then any minimizer for S(B(0, ρ), p, q) with q < q0 is radial and a multiple of ϕ1 . (3) Let 1 < q < p. Then the solution of (7.3) is unique and it is a multiple of ϕ1 . In particular, any extremal is a multiple of ϕ1 . Radial symmetry is lost when ρ or q is sufficiently large. Define the function Q(ρ) =

1 λ1 (ρ)p/(p−1)



λ1 (ρ) 1 − (N − 1) + 1. ρ

We have the following theorem. T HEOREM 7.3. Let 1 < p < ∞ be fixed. (1) Let ρ > 0. If q > Q(ρ) then there is no radial minimizer for S(B(0, ρ), p, q). (2) Let p < q < p(N − 1)/(N − p). There exists R(q) such that for any ρ > R(q) there is no radial minimizer for S(B(0, ρ), p, q).

Nonlinear boundary conditions

363

The main ideas of the proofs of these theorems are as follows. The proof of uniqueness of the solution of (7.3) for 1 < q < p follows from Picone’s identity and the weak formulation. In fact, assume that there exist two positive solutions u and v, then up 0 |∇u| dx − |∇v| ∇v∇ p−1 dx v B(0,ρ) B(0,ρ)    up =− up dx + uq dσ + v p−1 p−1 dx v B(0,ρ) ∂B(0,ρ) B(0,ρ)  p u − v q−1 p−1 dσ v ∂B(0,ρ)   = uq dσ − v q−p up dσ 

p

∂B(0,ρ)

=



∂B(0,ρ)



p−2



∂B(0,ρ)

 up uq−p − v q−p dσ.

(7.5)

We can interchange the roles of u and v in (7.5) and finally obtain 0



∂B(0,ρ)

p   u − v p uq−p − v q−p dσ.

Since q < p, the integrand is nonpositive and hence u = v on ∂B(0, ρ). The uniqueness follows from the uniqueness of the Dirichlet problem for
p u = up−1 . The function Q(ρ) appears when one try to prove that the extremals are not radial p−1 for large q. In fact, let u0 denote the positive radial solution of
p u0 = u0 in RN N normalized such that u0 = 1 on ∂B(0, ρ). For any t ∈ R and x ∈ R , denote by x t = (x1 − t, x2 , . . . , xN ) and consider the function Φ(t) =



p t t p B(0,ρ) |∇u0 (x )| + u0 (x ) dx  . q ( ∂B(0,ρ) u0 (x t ) dσ )p/q

After some calculations, see [92] for the details, we get Φ ′ (0) = 0 and λ1 (ρ) p/(p−1) Φ (0) = C 1 − (N − 1) . − (q − 1)λ1 (ρ) ρ ′′



Hence, when q > Q(ρ) we have Φ ′′ (0) < 0 and hence t = 0 is a local maximum for Φ. Therefore u0 is not a minimizer.

364

J.D. Rossi

One remarkable fact concerning the first eigenvalue λ1 (ρ) of (7.4) is that it verifies the differential equation λ′1 (ρ) = 1 − (p − 1)λ1 (ρ)p/(p−1) − (N − 1)

λ1 (ρ) , ρ

with the initial condition λ1 (0) = 1. The proof of this fact is as follows, see [92] for details. p−1 Let u0 denote the positive radial solution of
p u0 = u0 in RN normalized such that u0 (0) = 1. For any ρ > 0 the first eigenfunction of (7.4) is given by the restriction of u0 to B(0, ρ). From the boundary condition we get λ1 (ρ) =

u′0 (ρ)p−1 . u0 (ρ)p−1

Differentiating with respect to ρ and using the equation verified by u0 , 

N −1  ′  u (ρ)p−2 u′ (ρ) ′ = ρ N −1 up−1 (ρ), ρ 0 0 0

we get the desired differential equation

λ′1 (ρ) = 1 − (p − 1)λ1 (ρ)p/(p−1) − (N − 1)

λ1 (ρ) . ρ

To obtain the initial condition λ1 (0) = 1 we just observe that λ1 (ρ) → 1 as ρ → 0, see Theorem 6.3. The technique of spherical symmetrization, also known as foliated Schwarz symmetrization, is well suited for the study of the symmetry properties of nonradial extremals for our problem. The definition of spherical symmetrization (see [90]) is as follows. Given a measurable set A ⊂ RN , the spherical symmetrization A∗ of A is defined as follows: For each r, take A ∩ ∂B(0, r) and replace it by the spherical cap of the same area and center reN . The union of these caps is A∗ . Let u be an extremal and let u∗ denote the spherical symmetrization of u with respect to the north pole, eN . It is well known that, for any ball B(0, ρ), we have  ∗ u 

 ∗ u 

W 1,p (B(0,ρ))

Lp (∂B(0,ρ))

 uW 1,p (B(0,ρ))

= uLq (∂B(0,ρ)) .

and (7.6)

Hence, u∗ is also a minimizer. Remark that the restriction of u∗ to any sphere centered at the origin and contained in B(0, ρ) is an increasing function of the geodesic distance from the north pole. This fact together with the maximum principle imply that u∗ concentrates at a single point on the boundary of B(0, ρ). For the special case when p = 2, it is shown in [48] that either the inequality in (7.6) is strict or u and u∗ coincide on every sphere up to a rotation. This implies that any minimizer is spherically symmetric.

Nonlinear boundary conditions

365

8. Behavior of the best Sobolev trace constant and extremals in domains with holes In this section, using results from [76], we study the best Sobolev trace constant corresponding to the embedding of W 1,p (Ω) into Lq (∂Ω) for functions that vanish on a fixed subset of Ω, that we will call A. We consider subcritical exponents 1  q < p∗ = p(N − 1)/(N − p) so that the immersion W 1,p (Ω) ֒→ Lq (∂Ω) is compact. To begin with, we consider a subset A ⊂ Ω with positive measure and define the best Sobolev trace constant associated with this set, that is,


(|∇u|p + |u|p ) dx 1,p : u ∈ W (Ω), u| ≡  0 and u| = 0 . (8.1) SA = inf Ω  ∂Ω A ( ∂Ω |u|q dS)p/q Since along this section we are interested in the dependence of this best constant on A, we have dropped the explicit dependence of S on (Ω, p, q). As a consequence of the compact immersion, there exist extremals for SA . An extremal for SA is a weak solution to  p−2 u = 0 in Ω \ A,   −
p u + |u| ∂u q−2 p−2 (8.2) u on ∂Ω, |∇u| ∂ν = λ|u|   u=0 in A,

where λ depends on the normalization of u. For instance, if uLq (∂Ω) = 1 then λ = SA . The existence of weak solutions to (8.2) when A = ∅ has been studied in Section 6, see [71]. Of special importance is the case q = p in which the equation and boundary condition in (8.2) have the same homogeneity. In this case, (8.2) can be considered as a nonlinear eigenvalue problem. This nonlinear eigenvalue problem, in the case A = ∅, has been already studied in this work (Section 6), see also [71] and [100]. So we are studying a nonlinear generalization of an eigenvalue problem by adding the restriction that the functions involved vanish on a subset A. Optimal design problems are usually formulated as problems of minimization of the energy, stored in the design under a prescribed loading. Solutions of these problems are unstable to perturbations of the loading. The stable optimal design problem is formulated as minimization of the stored energy of the project under the most unfavorable loading. This most dangerous loading is one that maximizes the stored energy over the class of admissible functions. The problem is reduced to minimization of Steklov eigenvalues. See [28]. In view of the above discussion, our first concern is to consider the following optimization problem: For a fixed 0 < a < |Ω|, find a set A0 of measure a that minimizes SA among all subsets A ⊂ Ω of measure a. We have that such a set exists and, in the case that Ω is a ball, that there exists an optimal set that is spherically symmetric in the sense of [90] (see Section 7 for the definition). Moreover, in the case p = 2, every optimal set is spherically symmetric. On the other hand, we also have that there does not exists a set A that maximizes SA . Also, sup SA = +∞ where the supremum is taken over all sets A of given measure a. We also get the continuity of SA with respect to the “hole” A. Optimization problems for eigenvalues of elliptic operators has been widely studied in the past, and is still an area of intensive research. In [48] the author studies an optimization

366

J.D. Rossi

problem for the second Neumann eigenvalue of the Laplacian with A ⊂ ∂Ω. Our approach to the optimization problem follows closely the one in [48]. Optimal design problems have been widely studied not only for eigenvalue problems. Now we state the main results of this section. Our first result is the sequential lower semicontinuity of SA . T HEOREM 8.1. Let An ⊂ Ω be sets of positive measure such that ∗

in L∞ (Ω),

χAn ⇀ χA0

where χA is the characteristic function of the set A. Then SA0  lim inf SAn , n→∞

where SA is given by (8.1). We remark that the continuity is not true in general. This semicontinuity result suggest that a minimizer for SA among sets A of fixed positive Lebesgue measure exists. However, there is a major difficulty here because of the fact that sets of prescribed positive Lebesgue measure are not compact with respect to the topology of Theorem 8.1. The result concerning the existence of an optimal design for the constant SA is as follows. T HEOREM 8.2. Given 0 < a < |Ω|, let us define S(a) :=

inf

A⊂Ω,|A|=a

SA .

Then, there exists a set A0 ⊂ Ω such that |A0 | = a and SA0 = S(a). On the other hand, there is no upper bound for S(a). Let 0 < a < |Ω|. Then sup A⊂Ω,|A|=a

SA = ∞.

Next we study symmetry properties of optimal sets A0 in the special case where Ω is a ball. To this end, we use the definition of spherical symmetrization (see [90] and Section 7). We have the following result. T HEOREM 8.3. Let Ω = B(0, 1) and 0 < a < |B(0, 1)|. Then, there exists an optimal hole of measure a which is spherically symmetric, that is, A∗ = A. Moreover, when p = 2 every optimal hole is spherically symmetric. Now we state the results that allow us to consider the case of measure zero, that is, |A| = 0. For simplicity we will consider closed sets A. When trying to give sense to a best Sobolev trace constant for functions that vanish in a set of zero Lebesgue measure a dif1,p  \ A), where ferent approach has to be made. We consider the space WA (Ω) = C0∞ (Ω

Nonlinear boundary conditions

367

1,p

the closure is taken in W 1,p norm. That is, WA (Ω) stands for the set of functions of the Sobolev space W 1,p (Ω) that can be approximated by smooth functions that vanish in a neighborhood of A. In this context the best Sobolev trace constant is defined as  p p Ω|∇u| + |u| dx SA = inf . q p/q 1,p 1,p u∈W (Ω)\W (Ω) ( ∂Ω |u| dS) A

0

In this case this problem only makes sense if A is a set with positive p-capacity (see (8.3)). More precisely, we have that SA = S∅ if and only if the p-capacity of A is zero. Note that S∅ is the usual Sobolev trace constant from W 1,p (Ω) onto Lq (∂Ω). Observe that the constants SA and SA need not be the same. First, we study when SA is equal to the usual Sobolev trace constant, that is, when SA = S∅ . For this purpose we recall the definition of p-capacity. For A ⊂ Ω, define Capp (A)
 = inf

Rn





  |∇φ|p dx  φ ∈ W 1,p RN ∩ C ∞ RN and A ⊂ {φ
1}◦ .

(8.3)

We have the following theorem.

1,p

T HEOREM 8.4. Let A ⊂ Ω. Then WA (Ω) = W 1,p (Ω) if and only if Capp (A) = 0. As a corollary we obtain: C OROLLARY 8.1. Capp (A) = 0 if and only if SA = S∅ . Next, we look for the dependence of SA under perturbations of A. We find that SA is continuous with respect to A in the topology given by the Hausdorff distance. T HEOREM 8.5. Let A, An ⊂ Ω be closed sets such that d(An , A) → 0 as n → ∞ where d(An , A) is the Hausdorff distance between An and A. Then |SAn − SA | → 0 when n → ∞ and if we denote by un an extremal for SAn normalized such that un Lq (∂Ω) = 1, there exists a subsequence unk such that limk→∞ unk = u, strongly in W 1,p (Ω), and u is an extremal for SA .

9. The Sobolev trace embedding with the critical exponent In this section we look at problems with the critical Sobolev trace exponent in the boundary condition. Existence result for elliptic problems with critical Sobolev exponents have deserved a great deal of attention since the pioneering work [22] and is an intensive area of research nowadays.

368

J.D. Rossi

Now we describe the results of [35]. The authors study the problem   
u + λu = 0 u=0   − ∂u = uN/(N −2) xN

in B + (0, 1), on ∂B + (0, 1) ∩ {xN > 0}, on ∂B + (0, 1) ∩ {xN = 0}.

(9.1)

Here B + (0, 1) stands for the half-ball {|x| < 1, xN > 0}. The following result parallels the one in [22] for Dirichlet boundary conditions. T HEOREM 9.1. Let µ1 be the first eigenvalue of the problem   
φ + µφ = 0 φ=0   − ∂φ = 0 xN

in B + (0, 1), on ∂B + (0, 1) ∩ {xN > 0}, on ∂B + (0, 1) ∩ {xN = 0}.

Then (1) if N
4 then a positive solution to problem (9.1) exists if and only if 0 < λ < µ1 ; (2) for N = 3 there is no positive solution of (9.1) for λ
µ1 = π2 while for π2 /4 < λ < π2 positive solutions exist. There is a value λ∗ ∈ (0, π2 /4) such that no positive solution exists for −∞ < λ < λ∗ . Remark that, as happens for the Dirichlet problem (see [22]), N = 3 is a critical dimension. The main reason to deal with the half-ball B + (0, 1) is the fact that solutions are 2,α C (B + (0, 1)) and have cylindrical symmetry. For the nonexistence proof the authors use a sharp Pohozaev identity. For the existence part they consider the minimization problem Aλ =

inf

u∈H 1 (Ω),u=0 on {xN =0}




∇u2L2 (Ω) − λu2L2 (Ω)

u2L2(N−1)/(N−2) (∂B + (0,1)∩{x

N >0})

,

and they prove that when 0 < λ < µ1 (N
4) or λ∗ < λ < µ1 (N = 3) there is the case when Aλ lies below a critical level. Hence the concentration–compactness results of [96,97] can be used to obtain the existence of a minimizer for Aλ that turns out to be a solution of (9.1).  In [3] it is studied another linear perturbation of the Laplace equation with a critical nonlinearity at the boundary, namely,  in Ω,  
u = 0 u=0 on ΓD ,   − ∂u = uN/(N −2) + λu on Γ . N ∂ν

(9.2)

Here we assume that the boundary of Ω splits into two parts ΓD and ΓN . If we denote ˜ on ΓN , by µ˜ 1 the first eigenvalue of the problem
φ = 0 with φ = 0 on ΓD and ∂φ ∂ν = µφ

Nonlinear boundary conditions

369

then problem (9.2) has a positive solution if and only if 0 < λ < µ˜ 1 , no matter the dimension. Hence, although (9.2) and (9.1) are linear perturbations of the same problem their behavior for N = 3 is very different. For the existence result instead of Aλ they consider minimizing


∇u2 − λu2L2 (Γ ) L2 (Ω) N 1 Bλ = inf : u ∈ H (Ω), u = 0 on ΓD , u2L2(N−1)/(N−2) (Γ ) N

and prove that the minimum is attained if and only if 0 < λ < µ˜ 1 . To explain the different behavior of problems (9.2) and (9.1) for N = 3, we observe that the linear perturbation taking place on an (N − 1)-dimensional manifold instead of an N -dimensional domain typically reduces the critical dimension by 1, which leads to the fact that there is no critical dimension for problem (9.2).  Concerning the symmetry properties of the extremals of the Sobolev trace constant, it is proved in [65] that if Ω is a ball of sufficiently small radius, then the extremals are radial functions, see Section 7. Also in [65] the authors use this result to prove that there exists a radial extremal for the immersion H 1 (B(0, ρ)) → L2∗ (∂B(0, ρ)) if the radius ρ is small enough. See also [3] for other geometric conditions that leads to existence of extremals in the case p = 2. Concerning the existence of extremals for the Sobolev trace theorem with the critical exponent in a general smooth bounded domain Ω, the main result of this section is the following theorem. T HEOREM 9.2. Let Ω be a bounded smooth domain in RN such that 1 |Ω| < , |∂Ω|p/p∗ K(N, p)

(9.3)

where K(N, p) is the trace constant for RN + . Then there exists an extremal for the immersion W 1,p (Ω) → Lp∗ (∂Ω). The proof of Theorem 9.2 uses the same approach as in [15], properly adapted to our new context, see [75] for the details. The other key ingredient in the proof is the result of [20] where the author compute the optimal constant K(N, p). See also [94] for a similar result in the case p = 2. R EMARK 9.1. Let Ω be any smooth bounded domain in RN and let Ωµ = µΩ = {µx | x ∈ Ω}, where µ > 0. We observe that when µ is small enough, precisely µ
s}, s > 0, be the counting function. By the definition of sj , one has that

 n s 2 , T ∗ T = n(s, T ),

(A.10)

and for any self-adjoint nonnegative operator, one has sj (T ) = µj (T ). The counting function satisfies the Weyl inequality, n(s1 + s2 , T1 + T2 )  n(s1 , T1 ) + n(s2 , T2 ).

(A.11)

508

G. Rozenblum and M. Melgaard

For 0 < p < ∞ the trace class (or Neumann–Schatten class) Sp is a set of T ∈ B∞ (H1 , H2 ) for which the following functional is finite p

T Sp (H1 ,H2 :=

 % &p sj (T ) = p j

∞

s p−1 n(s, T ) ds.

(A.12)

0

If p
1 then (A.12) defines a norm on Sp and a quasinorm for p < 1. Evidently, T ∈ Sq (H1 , H2 ) if and only if 

q/2 

< ∞. tr |T |q = tr T ∗ T p

For 0 < p < ∞ the class Sw (H1 , H2 ) ⊂ B∞ (H1 , H2 ) is the set of all compact operators T such that the following functional is finite: 41/p 3 . T Spw := sup s p n(s, T ) s>0

p

The functional  · Spw is a quasinorm. The classes Sw (H1 , H2 ) are not separable (if p p dim H1 = dim H2 = ∞); a separable subspace S0 ⊂ Sw is defined by 0 /  p S0 := T ∈ sp  lim s p n(s, T ) = 0 . s→+0

p

Note that Sp ⊂ S0 . For a compact self-adjoint operator T we set

 n± (s, T ) = N (λ, ∞); ±T .

(A.13)

For operators T = T ∗ ∈ B∞ the following functionals are introduced: ∆p(±) (T ) := lim sup s p n± (s, T ),

(A.14)

δp(±) (T ) := lim inf s p n± (s, T ),

(A.15)

s→∞

s→∞

so that 0  δp(±) (T )  ∆p(±) (T )  ∞. The functionals ∆p(±) , δp(±) are continuous in sp and p do not change if their argument changes by an operator of the class S0 . A.9. Glazman’s lemma The spectral theorem leads to the variational principle for the distribution function. There are many different formulations of this principle, the one most convenient for our purposes is called the Glazman lemma, which plays an important role in both qualitative and quantitative spectral analysis.

Schrödinger operators with singular potentials

509

For an operator with quadratic form t[u], the number of points of the spectrum, finite or infinite on some interval, is described by the dimensions of subspaces in D(t) where the ratio t[u]/uH = χ[u] is controlled. In particular, N (λ; T ), for the operator T defined by the quadratic form t, equals the maximum of the dimensions of subspaces where χ[u] < λ, or minimum of the co-dimensions of subspaces where χ[u]
λ. Construction of subspaces satisfying these inequalities immediately gives estimates for N (λ; T ) from below, resp. from above. Taken together, they, in particular, prove various versions of the Birman– Schwinger principle; see Section 6.2. L EMMA A.3. Let T be a lower (upper) semibounded self-adjoint operator and M ⊂ D(t) be a linear subset, dense in the T -norm. Then 

 N± (λ, T ) = max dim L ⊂ M: ± t[u] − λu2 < 0, u ∈ L \ {0} , 

 N± (λ, T ) = min codim L ⊂ D(t): ± t[u] − λu2
0, u ∈ L .

(A.16) (A.17)

Here codim L for a subspace L ⊂ D(t) denotes the minimal number of orthogonality conditions which determine L. Minus and plus signs correspond to upper, resp. lower, semibounded operators. An important consequence of the Glazman lemma is that the distribution function depends on the quadratic form in a monotone way. Let, for example, two operators T1 and T2 correspond to quadratic forms t1 [u], t2 [u], so that D(t1 ) ⊂ D(t2 ) and t1 [u]
t2 [u], u ∈ D(t1 ). Then in (A.16), the set of subspaces over which we maximize is larger for T2 than for T1 , and therefore, N (λ, T2 )
N (λ, T1 ) for any λ. We collect some properties of the distribution function for the Schrödinger operator HV = −
− V . The following lemma follows immediately from the Glazman lemma. L EMMA A.4. (i) If V1  V2 pointwise, then N (−E; HV2 )  N (−E; HV1 ) for all E
0. In particular, N (−E; HV )  N (−E; HV+ ). (ii) For all α ∈ [0, 1] and E
0, N (−E; HV )  N (−αE; H(V (x)−(1−α)E)+ ),

(A.18)

where (V (x) − (1 − α)E)+ denotes the positive part of the potential V (x) − (1 − α)E. Looking at the graph of N (−E; HV ) one sees that integrating N (−E; HV ) with respect to E yields (minus) the sum of the negative eigenvalues. More generally, since  ∂ N (−E; HV ) = − δ(E − Ej ), ∂E j

where {Ej }j denote the eigenvalues of HV , one has the following result [84,85].

510

G. Rozenblum and M. Melgaard

L EMMA A.5. Let γ > 0 and Sγ ,d (V ) := Sγ ,d (V ) = γ



∞



γ Ej 0. Let h and h0 be the corresponding quadratic forms, and let Q(h) = Q(h0 ) = K. Let Tε , where ε > 0, be the operator determined by the variational triple {K; h0 [u] + εu2K , g = h0 − h}. Then N (−ε; H ) = n+ (1, Tε ). If H0 is positive definite, then (A.21) is also valid for ε = 0.

(A.21)

Schrödinger operators with singular potentials

511

The operator Tε is traditionally called the Birman–Schwinger operator. The proof of Theorem A.6 can be reduced to comparing the formulae  N (−ε, H ) = max dim L ⊂ K: h[u] + εu2 < 0 ,  n+ (1, Tε ) = max dim L ⊂ K: g[u] > h0 [u] + εu2 ,

which follow directly from Lemma A.3. The assertion concerning the case ε = 0 can be extended to any H0
0, but the formulation becomes more involved (see [14]).

A.11. Asymptotic perturbation lemma Generally, one should expect that if one perturbs an operator with a weaker one, the main properties must not change. The lemma we give here (established first in [16]) assigns concrete meaning to this vague statement, as it concerns asymptotics of the spectrum. L EMMA A.7. Let K be a compact self-adjoint operator, and for some q > 0 and any ε > 0, K may be represented a sum, K = Kε + Ke′ , where 

lim sup n± t, Kε′ t q  ε.

lim n± (t, Kε )t q = c± (ε),

t→+0

t→+0

Then there exist limits limε→0 c± (ε) = c± and limt→+0 n± (t, K)t q = c± . P ROOF. Fix some δ > 0. The Weyl inequality (A.11) gives n+ (t, K)  n+ (t (1 − δ), Kε ) + n+ (tδ, Kε′ ). Passing to lim sup, we obtain (+)

c+ = lim sup n+ (t, K)t q  c+ (ε)(1 − δ)q + δ −q ε. t→0

On the other hand, applying Weyl inequality to Kε = K + (−Kε′ ), we obtain n+ (t, K)
(−) n+ (t (1 + δ), Kε ) − n− (tδ, Kε′ ). Passing here to lim inf, we get for c+ = lim inft→0 N+ (t, −q K)t



 (−) c+
lim n+ t (1 + δ), Kε t q − lim sup n− tδ, Kε′ t q t→0

t→0

−q


c+ (ε)(1 + δ)

−δ

−q

ε.

Thus (−)

(+)

c+ (ε)(1 + δ)−q − δ −q ε  c+  c+  c+ (ε)(1 − δ)q + δ −q ε. (−)

(−)

(A.22)

We set here δ = ε 1/(q+1) so that δ −q ε → 0. Then (A.22) gives c+ = c+ = lim c+ (ε). 

512

G. Rozenblum and M. Melgaard

The lemma plays a crucial role whenever one wants to establish an asymptotic formula. Here the general scheme is to prove the asymptotic formula first for some regular case (say, smooth and compactly supported potentials if one considers Schrödinger operators) and then use some uniform estimate (say, the CLR estimate) in combination with Lemma A.7 to extend the asymptotics to more singular cases.

References [1] C. Adam, B. Muratori and C. Nash, Multiple zero modes of the Dirac operator in three dimensions, Phys. Rev. D 62 (2000), 085026, 9 pp. [2] C. Adam, B. Muratori and C. Nash, Chern–Simons action for zero-modes supporting gauge fields in three dimensions, Phys. Rev. D 67 (2003), 087703, 3 pp. [3] R. Adami and R. Teta, On the Aharonov–Bohm Hamiltonian, Lett. Math. Phys. 43 (1998), 43–53. [4] D. Adams and L. Hedberg, Function Spaces and Potential Theory, Grundlehren Math. Wiss., Vol. 314, Springer-Verlag, Berlin (1996). [5] Y. Aharonov and D. Bohm, Significance of electromagnetic potentials in the quantum theory, Phys. Rev. 115 (1959), 485–491. [6] Y. Aharonov and A. Casher, Ground state of a spin- 12 charged particle in a two-dimensional magnetic field, Phys. Rev. A 19 (1979), 2461–2462. [7] M. Aizenman and E. Lieb, On semi-classical bounds for eigenvalues of Schrödinger operators, Phys. Lett. A 66 (1978), 427–429. [8] W. Arendt and A.V. Bukhvalov, Integral representations of resolvents and semigroups, Forum Math. 6 (1994), 111–135. [9] J. Avron, I. Herbst and B. Simon, Schrödinger operators with magnetic fields: I. General interactions, Duke Math. J. 45 (1978), 847–883. [10] A. Balinsky and W.D. Evans, On the number of zero modes of Pauli operators, J. Func. Anal. 179 (2001), 120–135. [11] R. Benguria and M. Loss, On a theorem of Laptev and Weidl, Math. Res. Lett. 7 (2000), 195–203. [12] Yu. Berezanski, Expansions in Eigenfunctions of Self-Adjoint Operators, Transl. Math. Monographs, Vol. 17, Amer. Math. Soc., Providence, RI (1968). [13] F. Berezin and M. Shubin, The Schrödinger Equation, Math. Appl. Soviet Ser., Vol. 66, Kluwer, Dordrecht (1991). [14] M.Sh. Birman, The spectrum of singular boundary problems, Mat. Sb. (N.S.) 55 (1961), 125–174 (in Russian); English transl.: Amer. Math. Soc. Transl. 53 (1966), 23–80. [15] M.Sh. Birman and V.V. Borzov, On the asymptotics of the discrete spectrum of some singular differential operators, Probl. Mat. Phys. 5 (1971), 24–38 (in Russian); English transl.: Topics Math. Phys. 5 (1972), 19–30. [16] M.Sh. Birman and M. Solomyak, On the leading term of the spectral asymptotics for non-smooth elliptic problems, Functional Anal. Appl. 4 (1971), 265–275. [17] M.Sh. Birman and M. Solomyak, Spectral Theory of Selfadjoint Operators in Hilbert Space, Reidel, Dordrecht (1987). [18] M.Sh. Birman and M. Solomyak, Schrödinger operator. Estimates for number of bound states as functiontheoretical problem, Spectral Theory of Operators, Novgorod, 1989, Amer. Math. Soc. Transl. Ser. 2, Vol. 150, Amer. Math. Soc., Providence, RI (1992), 1–54. [19] Ph. Blanchard and J. Stubbe, Bound states for Schrödinger Hamiltonians: Phase space methods and applications, Rev. Math. Phys. 8 (1996), 503–547. [20] M. Braverman, O. Milatovic and M. Shubin, Essential selfadjointness of Schrödinger type operators on manifolds, Russian Math. Surveys 57 (2002), 641–692. [21] L. Bugliaro, C. Fefferman, J. Fröhlich, G.M. Graf and J. Stubbe, A Lieb–Thirring bound for a magnetic Pauli Hamiltonian, Comm. Math. Phys. 187 (1997), 567–582.

Schrödinger operators with singular potentials

513

[22] L. Bugliaro, C. Fefferman and G.M. Graf, A Lieb–Thirring bound for a magnetic Pauli Hamiltonian. II, Rev. Mat. Iberoamericana 15 (1999), 593–619. [23] S. Chanillo and R.L. Wheeden, Lp -estimates for fractional integrals and Sobolev inequalities with applications to Schrödinger operators, Commun. Partial Differential Equations 10 (1985), 1077–1116. [24] I. Chavel, Isoperimetric Inequalities. Differential Geometric and Analytic Perspectives, Cambridge Tracts in Math., Vol. 145, Cambridge University Press, Cambridge (2001). [25] P.R. Chernoff, Essential self-adjointness of powers of generators of hyperbolic equations, J. Funct. Anal. 12 (1973), 401–414. [26] F. Cobos and T. Kühn, Lorentz–Schatten classes and pointwise domination of matrices, Canad. Math. Bull. 42 (1999), 162–168. [27] R. Courant and D. Hilbert, Methods of Mathematical Physics, Vol. 1, Interscience, New York (1953). [28] M. Cwikel, Weak type estimates for singular values and the number of bound states of Schrödinger operators, Ann. Math. 106 (1977), 93–100. [29] H.L. Cycon, R.G. Froese, W. Kirsch and B. Simon, Schrödinger Operators with Application to Quantum Mechanics and Global Geometry, Springer-Verlag, Berlin (1987). [30] R. De la Bretéche, Preuve de la conjecture de Lieb–Thirring dans le cas des potentials quadratiques strictement convexes, Ann. Inst. H. Poincaré Phys. Théor. 70 (1999), 369–380. [31] A. Dufresnoy, Un exemple de champ magnétique dans Rν , Duke Math. J. 50 (1983), 729–734. [32] D.E. Edmunds and W.D. Evans, Spectral Theory and Differential Operators, Clarendon Press/Oxford University Press, New York (1987). [33] D. Elton, The local structure of zero modes producing magnetic potentials, Comm. Math. Phys. 229 (2002), 121–139. [34] L. Erdös, Magnetic Lieb–Thirring inequalities, Comm. Math. Phys. 170 (1995), 629–668. [35] L. Erdös, Dia- and paramagnetism for nonhomogeneous magnetic fields, J. Math. Phys. 38 (1997), 1289–1317. [36] L. Erdös and J.P. Solovej, Semiclassical eigenvalue estimates for the Pauli operator with strong nonhomogeneous magnetic fields. II. Leading order asymptotic estimates, Comm. Math. Phys. 188 (1997), 599–656. [37] L. Erdös and J.P. Solovej, Semiclassical eigenvalue estimates for the Pauli operator with strong nonhomogeneous magnetic fields. I. Nonasymptotic Lieb–Thirring-type estimate, Duke Math. J. 96 (1999), 127–173. [38] L. Erdös and J.P. Solovej, The kernel of Dirac operators on S 3 and R3 , Rev. Math. Phys. 13 (2001), 1247–1280. [39] L. Erdös and J.P. Solovej, Uniform Lieb–Thirring inequality for the three dimensional Pauli operator with a strong non-homogeneous magnetic field, Ann. Inst. H. Poincaré 5 (2004), 671–741. Available at http://xxx.lanl.gov/pdf/math-ph/0304017. [40] L. Erdös and J.P. Solovej, Magnetic Lieb–Thirring inequalities with optimal dependence on the field strength, J. Statist. Phys. 116 (2004), 475–506. Available at http://xxx.lanl.gov/pdf/math-ph/0306066. [41] L. Erdös and V. Vugalter, Pauli operator and Aharonov–Casher theorem for measure valued magnetic fields, Comm. Math. Phys. 225 (2002), 399–421. [42] C. Fefferman, The uncertainty principle, Bull. Amer. Math. Soc. 9 (1983), 129–206. [43] C. Fefferman, Stability of matter with magnetic fields, Partial Differential Equations and Their Applications, Toronto, ON, 1995, CRM Proc. Lecture Notes, Vol. 12, Amer. Math. Soc., Providence, RI (1997), 119–133. [44] K. Friedrichs, Spektraltheorie halbbeschränkter Operatoren und Anwendung auf die Spektraltheorie von Differenzialoperatorer, Math. Ann. 109 (1934), 465–487, 685–713. [45] M. Fry, Nonperturbative behavior of 3-dimensional fermionic determinants, paramagnetism of charged Fermions, Phys. Rev. D 54 (1996), 6444–6452. [46] M. Fry, Paramagnetism, zero modes and mass singularities in QED 1 + 1, 2 + 1, and 3 + 1 dimensions, Phys. Rev. D 55 (1997), 968–972. [47] M. Fry, Fermion determinants, Internat. J. Modern Phys. A 17 (2002), 936–945. [48] V. Geyler and E. Grishanov, Zero modes in a periodic lattice of Aharonov–Bohm solenoids, JETP Lett. 75 (2002), 354–356.

514

G. Rozenblum and M. Melgaard

[49] V. Glaser, H. Grosse and A. Martin, Bounds on the number of eigenvalues of the Schrödinger operator, Comm. Math. Phys. 59 (1978), 197–212. [50] A. Grigor’yan and S.-T. Yau, Isoperimetric properties of higher eigenvalues of elliptic operators, Amer. J. Math. 125 (2003), 893–940. [51] B. Helffer, J. Nourrigat and X. Wang, Sur le spectre l’équation de Dirac (dans R2 ou R3 ) avec champ magnétique, Ann. Sci. École Norm. Sup. 22 (1989), 515–533. [52] B. Helffer and D. Robert, Riesz means of bounded states and semi-classical limit connected with a Lieb– Thirring conjecture, I, J. Asymptot. Anal. 3 (1990), 91–103; II, Ann. Inst. H. Poincaré 53 (1990), 139–147. [53] H. Hess, R. Schrader and D.A. Uhlenbrock, Domination of semigroups and generalization of Kato’s inequality, Duke Math. J. 44 (1977), 893–904. [54] P.D. Hislop and I.M. Sigal, Introduction to Spectral Theory. With Applications to Schrödinger Operators, Appl. Math. Sci., Vol. 113, Springer-Verlag, New York (1996). [55] D. Hundertmark, On the number of bound states for Schrödinger operators with operator-valued potentials, Ark. Mat. 40 (2002), 73–87. [56] D. Hundertmark, A. Laptev and T. Weidl, New bounds on the Lieb–Thirring constant, Invent. Math. 140 (2000), 693–704. [57] D. Hundertmark, E.H. Lieb and L.E. Thomas, A sharp bound for an eigenvalue moment of the onedimensional Schrödinger operator, Adv. Theor. Math. Phys. 2 (1998), 719–731. [58] V. Ivrii, Microlocal Analysis and Precise Spectral Asymptotics, Springer Monogr. Math., Springer-Verlag, Berlin (1998). [59] A. Iwatsuka, The essential spectrum of two-dimensional Schrödinger operator with perturbed constant magnetic field, J. Math. Kyoto Univ. 23 (1983), 475–480. [60] L.V. Kantorovich and G.P. Akilov, Functional Analysis, 2nd Rev. Edition, Nauka, Moscow (1977) (in Russian); English transl.: Pergamon Press, Oxford (1982). [61] T. Kato, Fundamental properties of Hamiltonian operators of Schrödinger type, Trans. Amer. Math. Soc. 70 (1951), 195–211. [62] T. Kato, Schrödinger operators with singular potentials, Proc. Internat. Symp. Partial Differential Equations and the Geometry of Normed Linear Spaces, Jerusalem, 1972, Israel J. Math. 13 (1972), 135–148. [63] T. Kato, Perturbation Theory for Linear Operators, Reprint of the 1980 Edition, Classics in Mathematics, Springer-Verlag, Berlin (1995). [64] T. Kato and K. Masuda, Trotter’s product formula for nonlinear semigroups generated by the subdifferentials of convex functionals, J. Math. Soc. Japan 30 (1978), 169–178. [65] V. Kondratiev, V. Maz’ya and M. Shubin, Discreteness of spectrum and strict positivity criteria for magnetic Schrödinger operators, Comm. Partial Differential Equations 29 (2004), 489–521. [66] V. Kondratiev and M. Shubin, Discreteness of spectrum for the magnetic Schrödinger operators, Comm. Partial Differential Equations 27 (2002), 477–525. [67] A. Laptev, Dirichlet and Neumann eigenvalue problems on domains in Euclidean spaces, J. Funct. Anal. 151 (1997), 531–545. [68] A. Laptev, On the Lieb–Thirring conjecture for a class of potentials, Oper. Theory Adv. Appl. 110 (1999), 227–234. [69] A. Laptev and T. Weidl, Sharp Lieb–Thirring inequalities in high dimensions, Acta Math. 184 (2000), 87–111. [70] A. Laptev and T. Weidl, Recent results on Lieb–Thirring inequalities, Journées Équations aux Dérivées Partielles, La Chapelle sur Erdre, 2000, Exp. 20, Univ. Nantes, Nantes (2000). [71] H. Leinfelder, Gauge invariance of Schrödinger operators and related spectral properties, J. Operator Theory 9 (1983), 163–179. [72] H. Leinfelder and C.G. Simader, Schrödinger operators with singular magnetic vector potentials, Math. Z. 176 (1981), 1–19. [73] H. Leschke, R. Ruder and S. Warzel, Simple diamagnetic monotonicities for Schrödinger operators with inhomogeneous magnetic fields of constant direction, J. Phys. A 35 (2002), 5701–5709. [74] D. Levin and M. Solomyak, The Rozenblum–Lieb–Cwikel inequality for Markov generators, J. Anal. Math. 71 (1997), 173–193. [75] P. Li and S.-T. Yau, On the Schrödinger equation and the eigenvalue problem, Comm. Math. Phys. 88 (1983), 309–318.

Schrödinger operators with singular potentials

515

[76] E. Lieb, Bounds on the eigenvalues of the Laplace and Schrödinger operators, Bull. Amer. Math. Soc. 82 (1976), 751–753. [77] E. Lieb, The number of bound states of one-body Schrödinger operators and the Weyl problem, Proc. Amer. Math. Soc. Symp. in Pure Math., Vol. 36, R. Ossermann and A. Weinstein, eds, Amer. Math. Soc., Providence, RI (1980), 241–252. [78] E. Lieb, Kinetic energy bounds and their applications to the stability of matter, Proc. Nordic Summer School in Math., Lecture Notes in Phys., Vol. 354, Springer, Berlin (1989), 371–382. [79] E.H. Lieb and M. Loss, Analysis, 2nd Edition, Amer. Math. Soc., Providence, RI (2001). [80] E.H. Lieb, M. Loss and J.P. Solovej, Stability of matter in magnetic fields, Phys. Rev. Lett. 75 (1995), 985–989. [81] E.H. Lieb, J.P. Solovej and J. Yngvason, Asymptotics of heavy atoms in high magnetic fields: I. Lowest Landau band region, Comm. Pure Appl. Math. 47 (1994), 513–591. [82] E.H. Lieb, J.P. Solovej and J. Yngvason, Asymptotics of heavy atoms in high magnetic fields: II. Semiclassical regions, Comm. Math. Phys. 161 (1994), 77–124. [83] E.H. Lieb, J.P. Solovej and J. Yngvason, The ground states of large quantum dots in magnetic fields, Phys. Rev. B 51 (1995), 10646–10665. [84] E. Lieb and W. Thirring, Bound for the kinetic energy of fermions which proves the stability of matter, Phys. Rev. Lett. 35 (1975), 687–689; 35 (1975), 1116 (Erratum). [85] E. Lieb and W. Thirring, Inequalities for the moments of the eigenvalues of the Schrödinger Hamiltonian and their relation to Sobolev inequalities, Studies in Mathematical Physics: Essays in Honor of Valentine Bargmann, E. Lieb, B. Simon and A. Wightman, eds, Princeton University Press, Princeton, NJ (1976), 269–303. [86] M. Loss and B. Thaller, Optimal heat kernel estimates for Schrödinger operators with magnetic fields in two dimensions, Comm. Math. Phys. 186 (1997), 95–107. [87] M. Loss and H.-T. Yau, Stability of Coulomb systems with magnetic fields: III. Zero energy bound states of the Pauli operator, Comm. Math. Phys. 104 (1986), 283–290. [88] A. Martin, Bound states in the strong coupling limit, Helv. Phys. Acta 45 (1972), 140–148. [89] V.G. Maz’ya, On the theory of multi-dimensional Schrödinger operator, Izv. Akad. Nauk SSSR Ser. Mat. 28 (1964), 1145–1172 (in Russian). [90] V.G. Maz’ya, Closure in the metric of the generalized Dirichlet integral, Zap. Nauchn. Sem. Leningr. Otdel. Mat. Inst. Steklov (LOMI) 5 (1967), 192–195. [91] V.G. Maz’ya, Sobolev Spaces, Springer-Verlag, Berlin (1985). [92] V.G. Maz’ya and M. Otelbaev, Imbedding theorems and the spectrum of a certain pseudodifferential operator, Siberian. Math. J. 18 (1977), 1073–1087 (in Russian). [93] V.G. Maz’ya and T.O. Shaposhnikova, Theory of Multipliers in Spaces of Differentiable Functions, Monogr. Stud. Math., Vol. 23, Pitman, Boston (1985). [94] V.G. Maz’ya and M. Shubin, Discreteness of spectrum and positivity criteria for Schrödinger operators. Preprint: arXiv:math.SP/0305278. [95] V.G. Maz’ya and I.E. Verbitsky, Capacitary estimates for fractional integrals with application to partial differential equations and Sobolev multipliers, Ark. Mat. 33 (1995), 81–115. [96] V.G. Maz’ya and I.E. Verbitsky, The Schrödinger operator on the energy space: Boundedness and compactness criteria, Acta Math. 188 (2002), 263–302. [97] V.G. Maz’ya and I.E. Verbitsky, Infinitesimal form-boundedness and subordination criteria for the Schrödinger operator, http://front.math.ucdavis.edu/math.FA/0406050. [98] M. Melgaard, E.-M. Ouhabaz and G. Rozenblum, Negative discrete spectrum of perturbed multivortex Aharonov–Bohm Hamiltonian, Ann. Inst. H. Poincaré 5 (2004), 979–1012. [99] M. Melgaard and G. Rozenblum, Spectral estimates for magnetic operators, Math. Scannd. 79 (1996), 237–254. [100] M. Melgaard and G. Rozenblum, Eigenvalue asymptotics for weakly perturbed Dirac and Schrödinger operators with constant magnetic fields of full rank, Comm. Partial Differential Equations 28 (2003), 697–736. [101] A. Mohammed and G. Raikov, On the spectral theory of the Schrödinger operator with electromagnetic potential, Pseudodifferential Calculus and Mathematical Physics, Math. Topics., Vol. 5, Akademie-Verlag, Berlin (1994), 298–390.

516

G. Rozenblum and M. Melgaard

[102] A. Molchanov, On conditions for discreteness of the spectrum of self-adjoint differential equations of the second order, Trudy Moskov. Mat. Obshch. 2 (1953), 169–199 (in Russian). [103] I.M. Oleinik, On the essential self-adjointness of the Schrödinger operator on complete Riemannian manifold, Math. Notes 54 (1993), 934–939. [104] E.-M. Ouhabaz, Invariance of closed convex sets and domination criteria for semigroups, Potential Anal. 5 (1996), 611–625. [105] V.V. Peller, Hankel operators of class Sp and their applications, Mat. Sb. (N.S.) 113 (1980), 538–581. [106] L.D. Pitt, A compactness condition for linear operators on functional spaces, J. Oper. Theory 1 (1979), 49–54. [107] A.Ya. Povzner, On the expansion of arbitrary functions in characteristic functions of the operator −
u + cu, Mat. Sb. (N.S.) 32 (1953), 109–156 (in Russian). [108] G.D. Raikov, Eigenvalue asymptotics for the Schrödinger operator with homogeneous magnetic potential and decreasing electric potential, I. Behaviour near the essential spectrum tips, Comm. Partial Differential Equations 15 (1990), 407–434; Comm. Partial Differential Equations 18 (1993), 1977–1979 (Errata). [109] G.D. Raikov, Semiclassical and weak magnetic field asymptotics for the Schrödinger operator with electromagnetic potential, Ann. Inst. H. Poincaré Phys. Theor. 61 (1994), 163–188. [110] G.D. Raikov, Spectral asymptotics for the perturbed 2D Pauli operator with oscillating magnetic fields. I. Non-zero mean value of the magnetic field, Markov Process. Related Fields 9 (2003), 775–794. [111] G.D. Raikov and S. Warzel, Quasi-classical versus non-classical spectral asymptotics for magnetic Schrödinger operators with decreasing electric potentials, Rev. Math. Phys. 14 (2002), 1051–1072. [112] M. Reed and B. Simon, Methods of Modern Mathematical Physics. II. Fourier Analysis, Self-adjointness, Academic Press, New York–London (1975). [113] M. Reed and B. Simon, Methods of Modern Mathematical Physics. IV. Analysis of Operators, Academic Press, New York–London (1978). [114] M. Reed and B. Simon, Methods of Modern Mathematical Physics. I. Functional Analysis, 2nd Edition, Academic Press, New York (1980). [115] F.S. Rofe-Beketov, Conditions for the self-adjointness of the Schrödinger operator, Mat. Zametki 8 (1970), 741–751 (in Russian); English transl.: Math. Notes 8 (1970), 888–894. [116] G. Rozenblum, Distribution of the discrete spectrum of singular operators, Dokl. Akad. Nauk SSSR 202 (1972), 1012–1015. [117] G. Rozenblum, Asymptotic behavior of the eigenvalues of the Schrödinger operator, Mat. Sb. (N.S.) 93 (1974), 347–367, 487 (in Russian). [118] G. Rozenblum, Domination of semigroups and estimates for eigenvalues, Algebra i Analiz 12 (2000), 158–177 (in Russian); English transl.: St. Petersburg Math. J. 12 (2001), 831–845. [119] G. Rozenblum, Eigenvalue analysis of elliptic operators, Topics Math. Anal., Springer (2005), to appear. [120] G. Rozenblum, M. Shubin and M. Solomyak, Spectral theory for differential operators, Partial Differential Equations, Vol. 7, Encyclopaedia of Mathematical Sciences, Vol. 64, Springer-Verlag, Berlin (1994), 1–272. [121] G. Rozenblum and M. Solomyak, The Cwikel–Lieb–Rozenblyum estimator for generators of positive semigroups and semigroups dominated by positive semigroups, Algebra i Analiz 9 (1997), 214–236 (in Russian); English transl.: St. Petersburg Math. J. 9 (1998), 1195–1211. [122] S. Ruijsenaars, The Aharonov–Bohm effect and scattering theory, Ann. Phys. 146 (1983), 1–34. [123] Yu. Safarov and D. Vassiliev, The Asymptotic Distribution of Eigenvalues of Partial Differential Operators, Transl. Math. Monographs, Vol. 155, Amer. Math. Soc., Providence, RI (1997). [124] J. Schwinger, On the bound states of a given potential, Proc. Nat. Acad. Sci. U.S.A. 47 (1961), 122–129. [125] Z. Shen, On moments of negative eigenvalues for the Pauli operator, J. Differential Equations 149 (1998), 292–327. [126] Z. Shen, On moments of negative eigenvalues for the Pauli operator, J. Differential Equations 151 (1999), 420–455. [127] I. Shigekawa, Spectral properties of Schrödinger operators with magnetic fields for a spin 12 particle, J. Func. Anal. 101 (1991), 255–285. [128] M. Shubin, Spectral theory of elliptic operators on noncompact manifolds, Astérisque 207 (1992), 35–108. [129] M. Shubin, Private communication (2003).

Schrödinger operators with singular potentials

517

[130] B. Simon, On the growth of the number of bound states which increase in potential strength, J. Math. Phys. 10 (1969), 1123–1126. [131] B. Simon, Quantum Mechanics for Hamiltonians Defined as Quadratic Forms, Princeton University Press, Princeton, NJ (1971). [132] B. Simon, The bound states of weakly coupled Schrödinger operators in one and two dimensions, Ann. Phys. 97 (1976), 276–288. [133] B. Simon, Universal diamagnetism of spinless Boson systems, Phys. Rev. Lett. 36 (1976), 1083–1084. [134] B. Simon, An abstract Kato’s inequality for generators of positivity preserving semigroups, Indiana Univ. Math. J. 26 (1977), 1067–1073. [135] B. Simon, Kato’s inequality and the comparison of semigroups, J. Funct. Anal. 32 (1979), 97–101. [136] B. Simon, Maximal and minimal Schrödinger forms, J. Oper. Theory 1 (1979), 37–47. [137] B. Simon, Trace Ideals and Their Applications, Cambridge Univivesity Press, Cambridge (1979). [138] B. Simon, Functional Integration and Quantum Physics, Academic Press, New York (1979). [139] A.V. Sobolev, Asymptotic behaviour of the energy levels of a quantum particle in a homogeneous magnetic field, perturbed by a decreasing electric field. I, Probl. Mat. Anal. 9 (1984), 67–84 (in Russian); English transl.: J. Soviet Math. 35 (1986), 2201–2212. [140] A.V. Sobolev, Asymptotic behavior of energy levels of a quantum particle in a homogeneous magnetic field perturbed by a decreasing electric field. II, Probl. Mat. Fiz. 11 (1986), 232–248 (in Russian). [141] A.V. Sobolev, Asymptotics of the discrete spectrum of the Schrödinger operator in an electric and a homogeneous magnetic field, Zap. Nauchn. Sem. Leningrad. Otdel. Mat. Inst. Steklov (LOMI) 182 (1990), 131–141 (in Russian); English transl.: J. Soviet Math. 62 (1992), 2807–2814. [142] A.V. Sobolev , On the Lieb–Thirring estimates for the Pauli operator, Duke Math. J. 82 (1996), 607–635. [143] A.V. Sobolev, Lieb–Thirring inequalities for the Pauli operator in three dimensions, Quasiclassical Methods, Minneapolis, MN, 1995, IMA Vol. Math. Appl., Vol. 95, Springer-Verlag, New York (1997), 155–188. [144] A.V. Sobolev, Quasiclassical asymptotics for the Pauli operator, Comm. Math. Phys. 194 (1998), 109–134. [145] K. Tachizawa, On the moments of the negative eigenvalues of elliptic operators, J. Fourier Anal. Appl. 8 (2002), 233–244. [146] H. Tamura, Asymptotic distribution of eigenvalues for Schrödinger operators with homogeneous magnetic fields, Osaka J. Math. 25 (1988), 633–647. [147] H. Tamura, Asymptotic distribution of eigenvalues for Schrödinger operators with homogeneous magnetic fields, II, Osaka J. Math. 26 (1989), 119–137. [148] H. Tamura, Resolvent convergence in norm for Dirac operator with Aharonov–Bohm field, J. Math. Phys. 44 (2003), 2967–2993. [149] B. Thaller, The Dirac Equation, Texts Monographs Phys., Springer-Verlag, Berlin (1992). [150] S. Thangavelu, Lectures on Hermite and Laguerre Expansions, Math. Notes, Vol. 42, Princeton University Press, Princeton, NJ (1993). [151] N. Ural’ceva, The nonselfadjointness in L2 (Rn ) of elliptic operators with rapidly increasing coefficients, Zap. Nauchn. Sem. Leningr. Otdel. Mat. Inst. Steklov (LOMI) 14 (1969), 288–194 (in Russian). [152] I.E. Verbitsky, Nonlinear potentials and trace inequalities, The Maz’ya Anniversary Collection, Oper. Theory Adv. Appl. 110 (1999), 323–343. [153] T. Weidl, On the Lieb–Thirring constants Lγ ,1 for γ
1/2, Comm. Math. Phys. 178 (1996), 135–146. [154] T. Weidl, Another look at Cwikel’s inequality, Differential Operators and Spectral Theory, Amer. Math. Soc. Transl. Ser. 2, Vol. 189, Amer. Math. Soc., Providence, RI (1999), 247–254. [155] J. Weidmann, Linear Operators in Hilbert Spaces, Springer-Verlag, New York–Berlin (1980).

This page intentionally left blank

CHAPTER 7

Multiplicity Techniques for Problems without Compactness Sergio Solimini Dipartimento di Matematica, Politecnico di Bari, 70125 Bari, Italy E-mail: [email protected]

Contents Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1. Statement of the problems and sketch of the existence results . . . . . . . . . 1.1. Elliptic problems at critical growth on a bounded domain . . . . . . . . 1.2. Elliptic problems at subcritical growth on the whole domain . . . . . . 1.3. Concentration–compactness tools . . . . . . . . . . . . . . . . . . . . . 1.4. Natural constraint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5. Existence of a nontrivial solution for the problem at critical growth . . . 1.6. Existence of a nontrivial solution for the problem on the whole domain 2. Approximating problems and compactness of balanced sequences . . . . . . 2.1. Avoiding concentration . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2. Avoiding escaping masses . . . . . . . . . . . . . . . . . . . . . . . . . 3. Decay estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1. Decay estimates near concentration points . . . . . . . . . . . . . . . . 3.2. Decay estimates at drift points . . . . . . . . . . . . . . . . . . . . . . . 4. Multiplicity results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1. Krasnoselskii genus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2. Genus of a symmetric set . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3. Min–max classes on the natural constraint . . . . . . . . . . . . . . . . 4.4. Min–max classes on the double natural constraint . . . . . . . . . . . . 4.5. An estimate on the Morse index . . . . . . . . . . . . . . . . . . . . . . 4.6. Multiple solutions to the problem at critical growth . . . . . . . . . . . 4.7. Multiple solutions to the problem on the whole domain . . . . . . . . . 5. Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1. Review on the assumptions and open problems . . . . . . . . . . . . . . 5.2. Estimates in lower dimension . . . . . . . . . . . . . . . . . . . . . . . Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

HANDBOOK OF DIFFERENTIAL EQUATIONS Stationary Partial Differential Equations, volume 2 Edited by M. Chipot and P. Quittner © 2005 Elsevier B.V. All rights reserved 519

. . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . .

521 522 522 524 527 536 539 540 541 541 546 553 554 564 570 571 572 578 581 584 587 590 593 593 594 598 598

Multiplicity techniques for problems without compactness

521

Introduction The aim of this survey relies on focusing some recent multiplicity results for nonlinear problems with a lack of compactness which could probably find new and different applications. The corresponding existence results are well known long before but a full understanding of the multiplicity problem has required more specific and recent techniques. Here we shall show those techniques by stressing on the geometric ideas underlying them. More precisely, two main problems will be addressed: 1. Elliptic problems at critical growth on a bounded domain. 2. Elliptic problems at subcritical growth on the whole domain. For both the above problems there is a lack of compactness which is due to the existence of extremely concentrated solutions in case 1 and to the existence of solutions whose centers of mass escape to infinity in case 2. In both cases the existence results are available thanks to suitable hypotheses on the linear term which make the compactness degeneracy increase the functional which is going to be minimized. Thus deviation from compactness is possible in principle but it is not advantageous, under different hypotheses on the lower-order terms one would easily show nonexistence results. This survey includes the results in [13,19,20,43], some parts of which are borrowed with minor modifications, and it is in large part devoted to a careful analysis of the two above problems following mainly [19] and [13]. For the common features shared by these problems, they well offer the opportunity to clarify and to focus the ideas and the techniques employed in those works. Moreover, we shall make the exposition self-contained as far as possible and organized in a heuristic way. Nevertheless, a good knowledge of the use of the variational and topological methods in nonlinear analysis and some familiarity with the study of nonlinear elliptic equations is required to the reader. The exposition begins with an introductory part in which we recall the known facts concerning the two already mentioned problems and, after introducing suitable concentration– compactness tools, briefly sketches the main ideas which lead to prove the existence of a nontrivial solution, recovering compactness thanks to suitable estimates on the energy levels. In Section 2 we shall prove the compactness theorems which show how, substituting the Palais–Smale sequences with the sequences of solutions of approximating problems, the concentration–compactness tools introduced in the first section lead to complete compactness results. Section 3 is devoted to the proof of some decay estimates, inspired by the analysis of some particular cases, employed in the previous section. Section 4 concerns the proof of the multiplicity theorems, which will be given after a preliminary part in which the notion of genus, together with a recent variant, is introduced and employed to construct the suitable min–max approach. Finally, in Section 5 some concluding remarks are stated and a result which falls out of the main hypotheses assumed here is shown. Further introductory indications are given at the beginning of each section. Notation. Throughout the paper we make use of the following notation: • Ω denotes a bounded subset of RN . • Lp (Ω), 1  p  +∞, Ω ⊆ RN , denotes a Lebesgue space, the norm in Lp (Ω) is denoted by  · p .

522

S. Solimini

• H01 (Ω), Ω ⊂ RN , denote the Sobolev space obtained as closure of C0∞ (Ω), with respect to the norm u =

'

Ω

|∇u|2 dx

(1/2

.

• H −1 (Ω), Ω ⊂ RN , denotes the dual spaces of H01 (Ω). • If u ∈ H01 (Ω), Ω ⊂ RN , and if there is no risk of ambiguity, we denote also by u its extension to RN made by setting u ≡ 0 on RN \ Ω. • S denotes the Sobolev constant, i.e., S = inf{∇u22 /u22∗ | u ∈ H01 (Ω), u = 0}. • We denote by λ1 < λ2  λ3  · · ·  λn  · · · the sequence of the eigenvalues of the Laplacian operator −
on H01 (Ω) and by ϕ1 , ϕ2 , . . . , ϕn , . . . the corresponding sequence of orthonormal eigenvectors, we set En = Span{ϕ1 , ϕ2 , . . . , ϕn }. For every real number c we shall say that a sequence (un )n∈N is a Palais–Smale (briefly PS) sequence for the functional I : H 1 → R at level c if the following two conditions hold: (1) I (un ) → c; (2) dI (un ) → 0 in H −1 , where dI is the Fréchet derivative of I . We shall briefly say that (un )n∈N is a PS sequence if there exists a level c ∈ R such that (un )n∈N is a PS sequence at level c.

1. Statement of the problems and sketch of the existence results 1.1. Elliptic problems at critical growth on a bounded domain Let us consider the critical growth problem


∗ −2

−
u = |u|2 u=0

u + λu in Ω, on ∂Ω,

(CP)

where Ω is an open regular subset (without any shape condition) of RN (N
3), 2∗ = 2N /(N − 2) is the critical Sobolev exponent for the embedding of H01 (Ω) into Lp (Ω), and λ > 0. Several people have got involved with this problem (see [2,3]) and here, for the reader’s convenience, we summarize the main known results in the field. 1. If λ  0, the Pohozaev identity (see [37]) allows us to say that problem (CP) has, in general (for a star-shaped Ω), no nontrivial solution. 2. There exists a constant λ∗ ∈ [0, λ1 [ such that (CP) has a positive solution if λ ∈ ]λ∗ , λ1 [, where λ1 is the first eigenvalue of −
defined on H01 (Ω). When N
4 then λ∗ = 0 (see [11]). The existence of a nontrivial solution also for λ
λ1 has been subsequently proved in [12]. In the three-dimensional case and when Ω is a ball then λ∗ = λ1 /4. Moreover, by using also in this case a suitable version of Pohozaev identity we know that, for λ ∈ ]0, λ∗ [, (CP) has no radial solution (see [11,12]) but it is still unknown if there exist nonradial solutions (changing sign) to (CP).

Multiplicity techniques for problems without compactness

523

3. If N
4 and Ω is a ball, then for any λ > 0, (CP) has infinitely many changing sign solutions (which sometimes cannot be radial, as shown in [2,3]) which are built by using the particular symmetry of the domain Ω (see [24]). 4. If N
7 and Ω is a ball, then for each λ > 0, (CP) has infinitely many changing signradial solutions, see [39] and a previous paper by Cerami, Solimini and Struwe [14], where it is also shown that for N
6, (CP) has at least two (pairs of ) solutions on any smooth bounded domain. 5. When 4  N  6 and Ω is a ball there exists a constant λ∗ > 0 such that (CP) has no changing sign-radial solution if λ ∈ ]0, λ∗ [. So the bound N
7 in the previous result cannot be removed (see [2,3]). 6. In [19] the question about existence of infinitely many solutions to problem (CP), for any bounded smooth domain Ω ⊂ RN in the case N
7, is affirmatively answered. Furthermore, by the above mentioned result in [2,3], the compactness arguments, which can be also employed in the radial case, cannot be extended to lower values of N . 7. Finally, in [20] it is shown that, for λ ∈ ]0, λ1 [ and N
4, problem (CP) has at least N 2 + 1 (pairs of ) solutions (N + 1 for λ close enough to 0) improving thus the result in [14]. Such result has been extended in [15] to the case λ
λ1 . The main difficulty in dealing with problem (CP) is the existence of noncompact Palais– Smale sequences (PS sequences) of the corresponding functional Iλ (u) =

1 2



Ω

|∇u|2 −

λ 2



Ω

|u|2 −

1 2∗



Ω

∗

|u|2 ,

(1.1)

defined on the Hilbert space H01 (Ω). The behavior of noncompact PS sequences has been studied in [45] which, roughly speaking, assures the existence of a subsequence approximated by its weak limit plus terms which tend to concentrate around a finite number of points (see Theorem 1.6). This result allows a precise description of the behavior of noncompact PS sequences of the functional Iλ and an estimate of their possible levels, suggesting the idea to look for good levels in order to get compactness. We shall employ this analysis in conjunction with other suitable compactness techniques to deal with the multiplicity problem. Then, following [19], we shall show as, in dimension N
7, every min–max admissible class produces precompact PS sequences. This will follow as a consequence of a uniform bound theorem stated for bounded sets U of solutions to


−
u = |u|p−2 u + λu in Ω, u=0 on ∂Ω,

(SP)

with p ∈ [2, 2∗ ]. This result will require suitable a priori estimates on some norms of the functions in U . Such estimates will be employed to the aim of finding a suitable control on the functions and on their derivatives and, finally, a local Pohozaev identity will allow the proof of the following uniform bound theorem. T HEOREM 1.1 (Uniform bound through concentration estimates). Let N
7 and U be a bounded set in H01 (Ω) whose elements are solutions, for a fixed λ > 0, to problems (SP),

524

S. Solimini

for p varying in [2, 2∗ ]. Then U is uniformly bounded, i.e., there exists a constant C > 0 such that   sup sup u(x)  C. u∈U x∈Ω

The above result is equivalent to a compactness property in H 1 (Ω) (see [10]) which allows uniform L∞ estimates in the case of a precompact set of solutions. Though there is a lack of compactness for PS sequences, we have compactness for the bounded sets of solutions. Thus the key idea, in the case of a critical growth, relies in using the variational methods to solve slightly subcritical problems, where the usual arguments based on PS sequences produce solutions, and then to pass to the limit on such a set of solutions. In the light of these considerations, it becomes evident that Theorem 1.1 plays a crucial role in this program of work, indeed its proof has involved the major difficulties. Furthermore, we shall show how this technique allows to apply classical min–max arguments to problem (CP) and to prove, in this way, the existence of infinitely many solutions, as it is stated in the following theorem. T HEOREM 1.2 (Infinitely many solutions to (CP) in large dimension). If N
7, then problem (CP) admits infinitely many solutions.

As we have just observed, analogous multiplicity results, like the existence of infinitely many radial solutions to (CP) when Ω is a ball, can be obtained from Theorem 1.1 in the same way as Theorem 1.2 and so the uniqueness result in [3], Theorem A, leads to the following remark. R EMARK 1.1. The restriction N
7 in Theorem 1.1 cannot be removed. Indeed, the theorem is false for N  6. Theorem 1.2 does not give any answer to the existence of infinitely many solutions to (CP) when N  6. In such a case it is only known the negative answer for radial solutions and the affirmative one for symmetric domains. Here we shall show through different techniques that when N
4, for λ ∈ ]0, λ1 [, problem (CP) has at least N2 + 1 (pairs of ) solutions (see [20]). 1.2. Elliptic problems at subcritical growth on the whole domain Let us consider the problem  −
u + a(x)u = |u|p−2 u in RN ,

 u ∈ H 1 RN ,

(P)

where N
2, p > 2 and p < 2N /(N − 2) when N > 2, and the potential a(x) is a continuous function, positive in RN , except at most a bounded set, satisfying suitable decay assumptions. We do not impose any symmetry property to a(x).

Multiplicity techniques for problems without compactness

525

Problems like (P) naturally arise in various branches of Mathematical Physics, indeed the solutions to (P) can be seen as solitary waves (stationary states) in nonlinear equations of the Klein–Gordon or Schrödinger type, moreover, they present specific mathematical difficulties that make them challenging to the researchers. The solutions to problem (P) can be searched as critical points of the energy functional I : H 1 (RN ) → R defined by I (u) =

1 2



 1 |∇u|2 + a(x)u2 dx − N p R



RN

|u|p dx.

(1.2)

The usual variational methods, that allow to prove the existence of infinitely many solutions to (P) in a bounded domain, cannot be straightly applied to I . Indeed, the embedding j : H 1 (RN ) → Lp (RN ) is continuous but not compact, therefore the basic Palais–Smale condition is not satisfied by I at all the energy levels. This difficulty can be avoided when a(x) enjoys some symmetry. Indeed, the first known results have been obtained considering a(x) = a(|x|) or even a(x) = a∞ ∈ R+ \ {0} (see [8,9,16,17,35,42]). In this case, the restriction of I to Hr1 (RN ), the subspace of H 1 (RN ) consisting of spherically symmetric functions, restores compactness, because the embedding of Hr1 (RN ) into Lp (RN ) is compact. So, the existence of a positive solution to (P) can be proved either by using mountain pass theorem or by minimization on a natural constraint, while the existence of infinitely many solutions follows by standard minimax arguments. Moreover, it is worth recalling that, still under the assumption a(x) = a(|x|), one can also find the existence of infinitely many nonradial changing sign solutions, breaking the radial symmetry of the equation (see [6] and reference therein). When a(x) does not enjoy any symmetry property, the problem becomes more difficult and even proving the existence of one positive solution is not a trivial matter. This situation requires a deeper understanding of the nature of the obstructions to the compactness and the use of more subtle tools. Most of the researches have been concerned with the case lim

|x|→+∞

a(x) = a∞ > 0

(1.3)

so that (P) can be related to the “problem at infinity”, −
u + a∞ u = |u|p−2 u in RN .

(P∞ )

A first answer to the existence question has been given proving that, in some cases, being true some inequalities relating (P) and (P∞ ), the concentration–compactness principle can be applied and (P) can be solved by minimization [28]. This is the case, for instance, when a(x) is a continuous function that, besides (1.3) and some decay assumptions, satisfies 0 < δ1  a(x)  a∞

∀x ∈ RN .

(1.4)

Subsequently, a careful analysis of the behavior of the Palais–Smale sequences (see [4,7]) has allowed to state that the compactness can be lost (in the sense that a PS sequence does not converge to a critical point) if and only if such a sequence breaks into a finite number

526

S. Solimini

of solutions to (P∞ ) which are centered at points which go to infinity. As a consequence, it has been possible to give an estimate of the energy levels in which the PS condition fails in terms of the energy of such masses and to face better some existence and multiplicity questions for (P). Indeed, the existence of a positive solution to (P) has been proved (see [4]) even when a ground state solution cannot exist, that is, for instance, when, besides (1.3) and suitably decay assumptions, the potential satisfies the condition a(x) > a∞ ∀x ∈ RN ; moreover, under conditions (1.3), (1.4) and a suitable decay at infinity, it has been shown the existence of a changing sign solution in addition to the positive one (see [33]). To conclude this brief review of known results, let us mention that there is some other work involving the use of variational methods to deal with standing waves of nonlinear Schrödinger equations. Some of these papers mainly deal with the existence of solutions to (P) using mountain pass and comparison arguments. See, e.g., [21,38] as well as the references therein. In particular, we point out that in [38] the existence of a positive and a negative solution is proved, provided (i)

inf a(x) > 0,

RN

(ii)

lim

|x|→+∞

a(x) = +∞,

(1.5)

while in [5] the existence of a third changing sign solution is shown. Some other papers study cases in which the potential a(x) possesses nondegenerate critical points and depends on a parameter, i.e., it appears like ah (x) = a(hx), and contain results of multiplicity of positive solutions under restrictions on the size of h (see [1,23,36] and for a(x) of a special form [34]). Finally we remind that, under assumptions of periodicity on a, (P) has been shown to posses infinitely many solutions (see [18]). Following [13], we shall assume the function a to satisfy the following conditions. (a1 ) a ∈ C 1 (RN , R); (a2 ) lim inf|x|→+∞ a(x) = a∞ > 0; (a3 ) ∂∂ax4 (x)eα|x| −→ +∞ ∀α > 0, where ∀x ∈ RN \ {0}, x4 = x/|x|; |x|→+∞

(a4 ) there exists a constant c¯ > 1 such that   ∇τ a(x)  c¯ ∂a (x) x ∂ x4

∀x ∈ RN : |x| > c, ¯

where ∇τx a(x) denotes the component of ∇a(x) lying in the hyperplane orthogonal to x4 and containing x. Therefore, in such a setting, we shall prove the next theorems. The ingredients of the proof recall the techniques and the estimates employed in the previous case for the problem at critical growth on a bounded domain. The analogous of the subcritical problem (SP) is, in this case, the same problem (P) on a bounded domain or, more specifically, on a ball Br centered in the origin, with homogeneous Dirichlet boundary conditions. Given r > 0, we shall therefore consider the approximating problem


−
u + a(x)u = |u|p−2 u in Br (0), u=0 on ∂Br (0).

(APr )

Multiplicity techniques for problems without compactness

527

The first theorem states a compactness property of the bounded sets of solutions to the approximating problems. T HEOREM 1.3. Assume that a(x) satisfies (a1 )–(a4 ). Let U ⊂ H 1 (RN ) be a bounded set consisting of solutions to (APr ) for some r > 0. Then U is a precompact subset of H 1 (RN ). From such compactness property we shall deduce the infinite multiplicity result. T HEOREM 1.4. If a(x) satisfies the assumptions (a1 )–(a4 ), then (P) has infinitely many solutions. Let us consider a sequence of balls in RN , Brn (0) = {x ∈ RN : |x| < rn }, rn −→ +∞, n→+∞ and the related problems (Pn ) = (APrn ) approaching (P). Since it is possible to prove that, for every n, (Pn ) possesses infinitely many solutions, obtained by constructing infinitely many critical levels for the related functionals as minimax on suitable classes of functions, it is a natural idea considering sequences {un } of solutions to (Pn ), corresponding to minimax classes of the same type, and then trying to pass to the limit. Clearly, once again, we need to prove that such sequences are precompact. Hence some additional tool is needed to control the situation. This is again a local Pohozaev-type inequality that, combined with some uniform decay estimates and integral bounds on any bounded sequence of solutions to (Pn ), allows to conclude that, under our assumptions, the lack of compactness due to translations cannot occur for such sequences because it is possible, in principle, but it is not convenient in order to minimize the functional.  where R EMARK 1.2. It is worth pointing out that, if in (P) we replace RN by RN \ Ω, N Ω is any bounded smooth open set in R , Theorem 1.4 is still true, because the arguments we shall develop still hold after very simple modification.

1.3. Concentration–compactness tools The presence of the critical exponent in the Sobolev embedding and the unbounded measure of the domain does not allow using the classical compactness techniques based on Rellich theorem but requires more fine tools as the ones studied by Lions in [28,29]. Compactness theorems are due to Struwe [45,46] for the problem at critical growth and to Benci and Cerami [7] for the problem on the whole domain. We shall follow the approach pursued in [42], where a sufficiently general statement including the two previous results is given. To this aim let us begin by introducing some terminology. We shall call scaling of center x0 and modulus σ the mapping ρ : x → x0 + σ (x − x0 ). In order to have always the possibility to compose two scalings, we shall include among them also the translations, which are the product of two scalings with inverse moduli and different centers. In such a case there is no center, the modulus is of course 1 and the function is determined by the translation vector. If (ρn )n∈N is a sequence of scalings we shall

528

S. Solimini

say that it is diverging by concentration if the corresponding sequence of the moduli diverges to +∞. We shall say that it is diverging by vanishing if the corresponding sequence of the moduli converges to zero and that it is diverging by translation if the corresponding sequence of the moduli is bounded and bounded away from zero and the corresponding sequence of the centers or of the translation vectors is diverging. We shall say that (ρn )n∈N is diverging if every subsequence admits a subsequence which is diverging by concentration or by vanishing or by translation. Two sequences of scalings (ρn )n∈N and (ρn′ )n∈N are said to be mutually diverging if the sequence ((ρn )−1 ◦ ρn′ )n∈N is diverging. If ρ is a scaling with modulus σ and u is a function defined on RN , for α ∈ R fixed, we shall refer to the function λα u ◦ ρ as the scaled function u by ρ and we shall denote it by the symbol ρ(u). Fixed 1  p < +∞, we shall take α = N/p in order to keep invariant the Lp norm and α = N/p ∗ , for p < N , to keep invariant the H 1,p norm. We shall transfer to the scalings of the functions the same terminology which we have introduced for the scalings of the variable. It turns out that a sequence of scalings (ρn )n∈N is diverging if for every u the sequence (ρn (u))n∈N weakly converges to zero or, equivalently, if there exists at least one function u = 0 such that (ρn (u))n∈N weakly converges to zero. In such a case, we can pass to a subsequence which is diverging by concentration or vanishing or translation. D EFINITION 1.1. Let U ⊂ H 1,p (RN ) be a bounded subset, we shall say that U has a bounded scale if, for every diverging sequence of scalings (ρn )n∈N and for every sequence (un )n∈N ⊂ U , the sequence of scaled functions (ρn (un ))n∈N weakly converges to zero ∗ in Lp (RN ). The following theorem has been proved in [42]. T HEOREM 1.5. Let (un )n∈N be a given bounded sequence of functions in H 1,p (RN ), with index p satisfying 1 < p < N . Then, replacing (un )n∈N with a suitable subsequence, we can find a sequence of functions (ϕi )i∈N belonging to H 1,p (RN ) and, in correspondence of any index i, we can find a sequence of scalings (ρni )i∈N in such a way that the sequence 1,p (RN ), uniformly with respect to n, and that the sequence (ρni (ϕ) i )i∈N is summable in H ∗ i (un − i∈N ρn (ϕ)i )n∈N converges to zero in Lp . Moreover, we have that, for any pair of j indexes i and j , the two corresponding sequences of scalings (ρni )i∈N and (ρn )i∈N are mutually diverging, that +∞  i=0

p

ϕi 1,p  M,

(1.6) p

where M is the limit of (un 1,p )n∈N , and that the sequence (un − verges to zero in H 1,p (RN ) if and only if (1.6) is an equality.



i i∈N ρn (ϕ)i )n∈N

con-

R EMARK 1.3. We notice that the above theorem still holds in the more general context of Lorentz spaces L(p ∗ , q) for q > p, it does not hold in the case q = p. Moreover,

Multiplicity techniques for problems without compactness

529

it is equivalent to state the compact embedding of bounded subsets of H 1,p (RN ) with a bounded scale into L(p ∗ , q) for q > p (see [42]). R EMARK 1.4. In the above theorem all the limits ϕi are the weak limits of (ρni )−1 (un ). Since for any two indexes i, j the corresponding sequences of scalings are mutually diverging, there exists at most one index i such that ρni admits the limit scaling. We shall denote by 0 such an index. Thus, it is not restrictive to assume ρn0 = id and so that ϕ0 is the weak limit of the sequence. We always reserve the index 0 to this aim, by taking ϕ0 as the weak limit of (un )n∈N even when ϕ0 = 0 and it does not need to be taken into account in Theorem 1.5. For i
1 we can suppose, by passing to a subsequence, that every sequence (ρni )n∈N is diverging by concentration or by vanishing or by translation. It is quite clear from the above theorem that the deviation from compactness for a PS sequence of the problem at critical growth on a bounded domain can be controlled by sequences of scalings only diverging by concentration. Whereas, for the subcritical problem on the whole domain the analogous phenomenon can be controlled in Lp by scalings only diverging by translation. D EFINITION 1.2. Let (un )n∈N ⊂ H 1,p (RN ) be a given sequence. We shall say that (un )n∈N is a fragmented sequence in H 1,p or, respectively, in Lq if there exists a finite number k
1 of functions ϕ0 , ϕ1 , . . . , ϕk belonging to H 1,p (RN ) and in correspondence of any index i
1, there exists (ρni )i∈N such that  ai sequence of mutually diverging scalings 1,p N the sequence (un − ϕ0 − i ρn (ϕ)i )n∈N converges to zero in H (R ) or, respectively, in Lq . D EFINITION 1.3. Let (un )n∈N be a fragmented sequence. In the case all of the ρni for i
1, are diverging by concentration, we shall say that the sequence is concentrating. If all of the ρi for i
1, are diverging by translation we shall say that the sequence is broken. Now we are in a position to apply the above results to the analysis of the two elliptic problems under consideration. Let us begin with problem (CP). 1.3.1. Concentrating sequences. Given σ > 0 and x¯ ∈ RN , let us consider the following scaled function  ∗ ρ(u) = uσ : x → σ N/2 u x¯ + σ (x − x) ¯ ,

where the choice of the exponent α = N/2∗ makes the scaling operation ρ keep constant the norms ∇uσ 2 and uσ 2∗ . In order to produce estimates on the values of solutions u to (SP), we observe that v = |u| (extended by zero out of Ω) solves 

∗

−
v  bv 2 −1 + A,

 v ∈ H 1 RN , v
0,

(EI)

530

S. Solimini

in the sense of distributions, where b is any coefficient greater than one and A = ∗ − inf(bs 2 −1 − s p−1 − λs) (taken for 1  p  2∗ , s > 0) is a constant which does not depend on u and p. Since b can be trivially normalized, we shall always take b = 1 in (EI). So most of the estimates employed for the solutions to (SP) will be derived for solutions to (EI) in H 1 (RN ) and this will let us free from caring about the sign of u or taking into account the domain Ω. D EFINITION 1.4. Let (un )n∈N be a given sequence. We shall say that (un )n∈N is • a controlled sequence if each un is a solution to (EI), • a balanced sequence if each un solves (SP) for some p ∈ [2, 2∗ ]. R EMARK 1.5. As we have already pointed out, the absolute value of every solution to (SP) (under an extension by zero out of Ω and multiplied by a constant) is also solution to (EI). Therefore any sequence consisting, term by term, of the absolute value of a balanced sequence is a controlled sequence. On the other side, when we shall deal with controlled sequences, we shall know that each term is positive and that Ω = RN . R EMARK 1.6. Let (un )n∈N be any PS sequence for Iλ . Then, if a sequence of functions (ϕi )i∈N is as in Theorem 1.5, for every i
1, ϕi solves (CP) in H 1 (RN ) with λ = 0 and, in particular, (EI) with A = 0. Since every solution to (EI) with A = 0 has the H01 norm greater or equal to a positive constant (see Remark 1.10), then from the above results we get the existence of a constant C > 0 such that ϕi H 1 > C for every i
1. Since the sequence of such norms is 0 summable, one can conclude that there are only finitely many ϕi . Therefore we are lead to recall the main result due to Struwe in [45] as a corollary of Theorem 1.5 and so we have the following statement which is appropriate to deal with the present situation. T HEOREM 1.6. Let (un )n∈N be a noncompact PS sequence. Then, by replacing (un )n∈N with a suitable subsequence, there exists a finite number k, depending on a bound M on un H 1 (namely k  MS −N/2 , where S is the Sobolev constant), of global solutions ϕi 0 to (CP) in H 1 (RN ) with λ = 0 with corresponding k sequences of mutually diverging scalings (ρni )n∈N with respective concentration points xni and diverging moduli σni (i.e., limn→+∞ σni = +∞) such that un −

k  i=1

ρni (ϕi ) → ϕ0

in H01 (Ω),

(1.7)

where ϕ0 is the weak limit of the sequence and solves (CP). P ROOF. Let (un )n∈N be a PS sequence and let vn ∈ H01 minimize the distance from k k i 1 N i i=0 ρn (ϕi ) in H (R ) for every n ∈ N. Then we have vn − i=0 ρn (ϕi ) → 0

Multiplicity techniques for problems without compactness

531

in H 1 (RN ). We notice that (vn )n∈N is a PS sequence, since all of the ϕi are solutions to (CP) in H 1 (RN ). Then, by an integration by parts, by denoting by εn a small term in the H −1 norm, we get 

RN

=

  ∇(un − vn )2 dx 

Ω

=− = =



  ∇(un − vn )2 dx



Ω



Ω

+

Ω


(un − vn )(un − vn ) dx

 ∗ ∗ |un |2 −2 un − |vn |2 −2 vn + λ(un − vn ) + εn (un − vn ) dx

 ∗ ∗ |un |2 −2 un − |vn |2 −2 vn + λ(un − vn ) (un − vn ) dx



Ω

εn (un − vn ) dx. ∗

∗

∗

∗′

Since un and vn are bounded in L2 , |un |2 −2 un and |vn |2 −2 vn are bounded in L2 . More∗ over, by Theorem 1.5 we know that un − vn → 0 in L2 , so we have the first term on the right-hand side converging to zero by duality. Furthermore, since un and vn are bounded in H 1 (RN ) and εn → 0 in H −1 we have that also the second term on the right-hand side converges to zero in H 1 and so the thesis is proved.  Theorem 1.6 says, in other terms, that from any noncompact PS sequence we can extract a concentrating sequence in H 1 . Given any concentrating sequence, we shall also consider the scalings ρni and the limit functions ϕi (which are not uniquely determined by Theorem 1.6) as also given. The next statement, which can be seen as a variant of Theorem 1.6, allows us to say that also from a noncompact balanced sequence (un )n∈N we can always extract a concentrating sequence, even if we do not know if (un )n∈N is a PS sequence. L EMMA 1.1. Let (un )n∈N be a noncompact bounded balanced sequence in H01 (Ω). Then from (un )n∈N we can extract a concentrating subsequence in H01 . P ROOF. Under a null extension of un to the whole of RN , we can use the structure theorem for bounded sequences Theorem 1.5. Then we shall show that for every i
1, |ϕi | solves (EI) with A = 0. Indeed, (ρni )−1 |un | → |ϕi | and |un | solves (EI). If we denote by νn the modulus of (ρni )−1 we have νn → 0. By scaling |un |, we have

−1

−1 ∗ (N +2)/2 −
ρni |un |  ρni |un |2 −1 + νn A,

(1.8)

532

S. Solimini

which, passing to the weak limit, gives (EI) for v = |ϕi | and A = 0. This implies, in particular, that ϕi 
S N/4 , where S is the Sobolev constant, see Remark 1.10, so we have an estimate on the number of the limits ϕi , i
1. Finally, we can show that (1.7) holds as in the proof of Theorem 1.6.  R EMARK 1.7. A more detailed argument, see [19], Lemma 6.2, shows that, while ϕ0 solves (CP) on Ω, every ϕi for i
1, is a solution to (CP) on the whole domain for λ = 0 multiplied by the number −N (N −2)/4(1−pn /2∗ )

µ = lim νn n


1.

From now on we shall denote by σni the modulus of the scaling ρni , so that for every given i we have σni → +∞ as n → +∞. For every i, j , through a selection argument, we can suppose that, for every n ∈ N, j i σn  σn (or vice versa), then we can order the indexes in such a way that, for every n ∈ N, it results σn1  σn2  · · ·  σnk . Therefore, with such an ordering, we have that ρn1 corresponds to a function concentrating in xn1 in the slowest way. For every n ∈ N, we set σn = σn1 and xn = xn1 . To the aim of establishing some local uniform estimates around the concentration points, we perform the following construction. In view of making estimates at a distance of the −1/2 from the concentration point xn , we need to exclude that, for a suitable conorder of σn stant c, someone of the functions ϕi , for i
1, could have a concentration point xni closer −1/2 (let us recall that xn corresponds to the function concentrating to ∂Bcσ −1/2 (xn ) than σn n in the slowest way), thus we argue as follows. For any n ∈ N, let us consider k concentric −1/2 and centered in xn . Since, by Theorem 1.6, the total number of annuli of width 7σn global solutions ϕi is k, we are sure that among those annuli there is at least one without any concentration point xni . Let A0n be that annulus. Since k  MS −N/2 does not depend  which by n, this procedure allows, passing to a subsequence, to choose a constant C,   7k + 1  7MS −N/2 + 1, and such that A0n = does not depend on n, such that 1  C   −1/2 (xn ). Then we set A1n = B  B(C+7)σ −1/2 (xn ) \ B −1/2 (xn ), −1/2 (xn ) \ B  Cσn (C+6)σn (C+1)σn n 2 3   An = B  −1/2 (xn ) \ B  −1/2 (xn ), An = B  −1/2 (xn ) \ B  −1/2 (xn ), get(C+5)σn

(C+2)σn

(C+4)σn −1/2

−1/2

(C+3)σn −1/2

−1/2

ting in this way four sequences of annuli, of width 7σn , 5σn , 3σn and σn , −1/2 i−1 i respectively, such that, for i = 1, 2, 3, An is the σn -neighborhood of An . So when i increases Ain gets thinner and we are going to establish finer estimates on it. When i = 0 we only know that A0n does not contain concentration points xni , we shall see in Section 2 how this rough estimate improves, in the case of a balanced sequence, for i = 1, 2, 3. When  we shall deal with a balanced sequence (un )n∈N , we shall assume to have fixed a constant C i as above and so we shall consider the four sequences An as also given and we shall call such sets safe regions of (un )n∈N .

Multiplicity techniques for problems without compactness

533

1.3.2. Broken sequences. We pass now to the analogous analysis for problem (P). Let us begin by considering some inequalities, related to (P), that will be useful in producing estimates on the solutions to the approximating problems:  p−1 in RN ,   −
u + a(x)u  u (EI1 ) u
0 in RN , 

  1 N u∈H R ,  p−1   −
u + a∞ u  u u
0 

  u ∈ H 1 RN .

in RN , in RN ,

We remark that if u is weak solution of  −
u + a(x)u = |u|p−2 u in Ω, u ∈ H01 (Ω),

(EI∞ )

(PΩ )

Ω ⊆ RN , then |u|, extended by 0 out of Ω, is a weak solution of (EI1 ). According to the previous definitions, we shall say that (un )n∈N is a balanced sequence if, for every n, un is a nontrivial weak solution to (APrn ), where (rn )n∈N , rn ∈ R+ , is any sequence so that rn > 0 and that it is a controlled sequence if, for each n, un is a nontrivial weak solution to (EI1 ). Once again we remark that to any balanced sequence (un )n∈N there corresponds a controlled sequence (vn )n∈N , where vn = |un | in Brn (0) and vn = 0 in RN \ Brn (0). R EMARK 1.8. We see that, given any balanced sequence (un )n∈N and a sequence of translation vectors (tn )n∈N , tn ∈ RN , |tn | −→ +∞, if n→+∞

u(·) = lim un (· − tn ) n→+∞

a.e. in RN ,

then |u| is a weak solution of (EI∞ ). A basic tool to face problems in unbounded domains has been the analysis of the PS sequences behavior and the information that in the framework of Theorem 1.5, when (1.3) is satisfied, a noncompact PS sequence differs from its weak limit by one or more sequences that, after suitable translations, go to a solution of (P∞ ) (see [7] and [4]). Here, since our aim relies in finding solutions to (P) that are limit of balanced sequences, we need to know how a noncompact bounded balanced sequence can look like. Moreover, since we want to analyze a rather general case by working with (a2 ) instead than (1.3), we cannot state the existence of a limit equation corresponding to (P). Nevertheless, by virtue of the previous considerations about the behavior of the balanced sequences and taking into account Remark 1.8, it is easy to realize that in our case the role of the limit problem can be played by (EI∞ ). The following lemma gives the necessary information leading to

534

S. Solimini

the conclusion that the set of norms of the solutions to (EI∞ ) is bounded from below by a positive constant. L EMMA 1.2. There exists a positive constant C0 > 0 such that for any nontrivial solution ϕ to (EI∞ ) ϕp
C0

(1.9)

holds. P ROOF. Let ϕ be a nontrivial solution to (EI∞ ), then ϕ satisfies p

∇ϕ22 + a∞ ϕ22  ϕp . By using Sobolev embedding theorem and by interpolating the Lp norm (taking into account that 2 < p < 2∗ ) we have

p p Sϕ22∗ + a∞ |ϕ|22  ϕp  ϕα2∗ ϕ21−α ,

where S denotes the best Sobolev constant and α ∈ (0, 1) is such that α/2∗ + (1 − α)/2 = 1/p. By applying Young inequality, we get ϕα2∗ ϕ21−α  αϕ2∗ + (1 − α)ϕ2 2 &1/2 % √ √ Sϕ2∗ + a∞ ϕ2  k1 1/2

,  k1 21/2 Sϕ22∗ + a∞ ϕ22

√ √ where k1 is chosen so that k1
max(α/ S, (1 − α)/ a∞ ). Hence,

p/2−1 Sϕ22∗ + a∞ ϕ22


1 k1 2p/2

and so we deduce, as desired, ϕp
C0 > 0, where C0 is a constant not depending on ϕ.  Taking advantage of Lemma 1.2 and by using the previous arguments, we can state the following assertions as corollaries of Theorem 1.5. Such results provide the desired picture of the controlled and therefore of the balanced sequences behavior. P ROPOSITION 1.1. Let a(x) satisfy (a1 ) and (a2 ). Let (un )n∈N be a noncompact controlled sequence bounded in H 1 (RN ). Then, there exists a subsequence (still denoted by un ) for which the following holds: there exist an integer k > 0, nontrivial solutions to (EI∞ ) ϕi ,

Multiplicity techniques for problems without compactness

535

1  i  k, sequences (tni )n∈N , 1  i  k, such that un −

k  i=1

 ϕi · − tni → ϕ0

 i t  −→ +∞, n n→+∞

 in Lp RN ,

 i t − tnj  −→ +∞, n n→+∞

(1.10) 1  i = j  k.

Now we are going to apply the same idea, used in dealing with the concentrating sequences, to the present case of diverging sequences. Specifically, we shall proceed in ordering the sequence of translation vectors and in constructing suitable regions on which we shall establish local uniform estimates. Thus, given any broken sequence (un )n∈N , we assume as given also the functions ϕi and the translation vectors tni (even if they are not uniquely determined) that appear in (1.10). Through a selection argument, we can suppose j that, for every n ∈ N, |tni |  |tn | (or vice versa), then we can order the indexes in such a way that, for every n ∈ N, it results |tn1 |  |tn2 |  · · ·  |tnk |. For every n ∈ N, we set tn = tn1 the basic sequence of translations. In view of constructing the safe regions of the space to associate to any broken sequence, we introduce the following terminology. D EFINITION 1.5. Let A ⊂ RN be a subset of RN and v ∈ RN a point v ∈ / A. We call cone of vertex v generated by A the smallest set containing A and positively homogeneous with respect to the vertex v, i.e., the set  w ∈ RN | w = v + λ(x − v), x ∈ A, λ ∈ R+ .

Let (un )n∈N be a broken sequence and let (tn )n∈N be the above defined sequence. In view of making estimates involving diverging sequences, we shall work in the safe regions connected with the basic sequence of translations in which we can deduce some a priori estimates which are not affected by the presence of other masses which are escaping to infinity. This time we shall not be concerned with annuli (so set differences of concentric balls) centered in xn but with set differences of coaxial cones with the axis parallel to tn . In order to avoid the other masses, we have to perform a similar argument to the one used for (CP). To this aim we proceed in constructing the following sequences of subsets of RN related to (tn )n∈N . For any n ∈ N, let us consider the cone Cn with vertex tn /2 and generated by a ball BRn (tn ). We begin by taking the cone C1,n generated by the ball B1,n = Brn (tn ), where  γˆ |tn | 1 1 rn = , (1.11) with 0 < γˆ < min , k 2 5 4(c¯ + 1) c¯ being the constant appearing in (a4 ). If ∂C1,n ∩ Brn /2 (tni ) = ∅ for all tni = tn , 1  i  k, we set Cn = C1,n and Rn = rn , otherwise we consider the larger cone C2,n having vertex tn /2 and generated by B2rn (tn ). Since |tn |  |tni |, 1  i  k, for any index i for which ∂C1,n ∩ Brn /2 (tni ) = ∅, we have Brn /2 (tni ) ⊂ C2,n , and we set Cn = C2,n if ∂C2,n does not touch any of the other balls Brn /2 (tni ), tni = tn . Otherwise we pass to the cone C3,n , having

536

S. Solimini

vertex tn /2, generated by B3rn (tn ) that surely contains the balls, of radius rn /2 centered at the points tni , touching ∂C2,n . We iterate this procedure and, since the number of the functions ϕi is k, we are sure that after at most k steps, we can associate to tn a cone Cn , having vertex tn /2, generated by a ball BRn (tn ), with γkˆ |t2n | = rn  Rn  krn = γˆ |t2n | , having the property that ∂Cn ∩ Brn /2 (tni ) = ∅, for any index i, 1  i  k, such that tni = tn . R EMARK 1.9. Let θn denote the width angle of the cone Cn . We emphasize that, since Rn = |t2n | tan θn , then  1 γˆ 1 .  tan θn  γˆ < min , 0< 2k 5 4(c¯ + 1) Now we introduce some tools which will be useful in dealing with problem (P). Let s ∈ R and n ∈ N, we consider the cones Cs,n = Cn − s t4n

(1.12)

and the regions around the boundary of Cn S2s,n = Cs,n \ C−s,n .

(1.13)

Lastly we set Sn = RN

k 8*

i=0

 Brn /2 tni .

(1.14)

1.4. Natural constraint The existence and the multiplicity results related to problems (CP) and (P) will be achieved by working with variational methods on the so-called natural constraint manifold, which is a subset V of H01 which contains all the critical points of a functional I and such that every constrained critical point of I on V is a critical point with respect to the whole space. We shall introduce this concept for the case of (CP), the other case is analogous. Let    1 λ 1 I (u) = |∇u|2 − |u|2 − |u|p (1.15) 2 Ω 2 Ω p Ω be the functional corresponding to (SP). We introduce the natural constraint V for the functional I as the manifold defined by  V = u | u = 0, ∇I (u) · u = 0 .

(1.16)

Multiplicity techniques for problems without compactness

Then the defining equation for the natural constraint is    |u|p = 0. Φ(u) = |∇u|2 − λ |u|2 − Ω

Ω

537

(1.17)

Ω

P ROPOSITION 1.2. For every u ∈ V, the following properties hold true: (i) I (u) > 0; (ii) I (u) > c > 0, if λ < λ1 ; (iii) I (u)
N1 S N/2 , if λ = 0 and p = 2∗ , where S is the Sobolev constant. P ROOF. By (1.17), a simple computation gives that, for every u ∈ V,      1 1 1 1 I (u) = − − |∇u|2 − λ |u|p > 0. |u|2 = 2 p 2 p Ω Ω Ω

(1.18)

So by (1.18), (i) follows. If λ < λ1 , then   |∇u|2 − λ |u|2
cu2H 1 (Ω)
cu2p Ω

Ω

and hence, by (1.17), p

u2p  cup ,

(1.19)

from which we have by the Sobolev embedding up
c. Finally, by combining this last inequality with (1.18) we have   1 1 − I (u) = |u|p
c 2 p Ω and so we get (ii). Moreover, when therefore, by the Sobolev inequality, S N/2 . Then, by (1.18), we finally get  1 1 ∗ I (u) = − ∗ u22∗
2 2

∗

λ = 0 and p = 2∗ , u ∈ V implies ∇u22 = u22∗ ∗ ∗ ∗ ∗ we have S  u22∗ −2 and so u22∗
S 2 /(2 −2) = 1 N/2 S N

from which (iii) follows.



R EMARK 1.10. The proof of (iii) also applies to positive solutions to (EI) when A  0, ∗ ∗ since we only use the inequality ∇u22  u22 , trivially implied by (EI). The main property of V is stated in the following proposition. P ROPOSITION 1.3. Let (un )n∈N ⊂ V be a constrained PS sequence for Iλ . Then (un )n∈N is a PS sequence for Iλ .

538

S. Solimini

P ROOF. Passing to a subsequence we can assume I (un ) → c
0. If c = 0, by (1.18) p we get up → 0, so u2 → 0 by Hölder inequality and finally uH 1 = (up + 2 λu2 ) → 0 by the constraint equation (1.17). So un → 0 and therefore (un )n∈N is a PS sequence. Thus let us assume c > 0. For every n ∈ N, we have ∇I (un ) = µn ∇Φ(un ) + rn , where rn is an infinitesimal term in H −1 (Ω) and µn is the Lagrange multiplier. Then, for every n ∈ N,

 ∇I (un ) = µn ∇ 2 I (un ) · un + ∇I (un ) + rn .

(1.20)

After an easy computation, multiplying both the sides of (1.20) by un and integrating, we get from (1.20)    2 2 p (1 − µn ) |∇un | − λ |un | − |un | = µn

Ω



2

Ω

|∇un | − λ



Ω

Ω

Ω

2

|un | − (p − 1)



Ω

p

|un |



+



Ω

rn · un ,

and, by taking into account the constraint equation (1.17), we have (p − 2)µn



Ω

|un |p =



Ω

rn · un .

Now, by estimating the term on the right-hand side of the previous equation and by (1.17), we obtain 

Ω

rn · un  rn H −1 (Ω) un H 1 (Ω)

1/2 p  rn H −1 (Ω) un p + λun 22 ,

where c is a positive constant. So by Hölder inequality,

−p/2 1−p  (p − 2)µn  crn H −1 (Ω) un p + un p . p

Since rn H −1 (Ω) → 0 and un p → c > 0, we get µn → 0 as n → ∞.

2p p−2 c

> 0, by virtue of (1.18) and the condition 

C OROLLARY 1.7. Let u ∈ V be a constrained critical point for Iλ . Then u is critical point for Iλ .

Multiplicity techniques for problems without compactness

539

1.5. Existence of a nontrivial solution for the problem at critical growth In this subsection we shall only sketch the main arguments involved in the existence theory in the case N
4 and λ < λ1 , we refer to [46] and the references therein for a more detailed treatment of the subject. On the other hand, the existence result is trivially implied by the multiplicity theorem which we are going to prove in Section 4 and which does not make use of this existence result. The only reason for which we are giving this sketch is to show the difference between a compactness argument based on a level estimate, which is enough to give the existence of a nontrivial solution and the compactness techniques required for the multiplicity results. We recall that for λ < λ1 , since zero is an isolated point for V, we also have infV Iλ > 0. To prove the existence of a nontrivial solution to the minimization problem for Iλ we shall assume to work with any minimizing sequence on V and then, by applying the results of Section 1.3 and after some estimates, we shall conclude that along any of such a sequence the functional cannot reach a level corresponding to a noncompact PS sequence and so, by recovering compactness, the standard variational methods allow to state the existence of a solution. C OROLLARY 1.8. Let (un )n∈N be a noncompact PS sequence for Iλ . Then Iλ (un ) → c
N −1 S N/2 as n → ∞. P ROOF . Let (un )n∈N be a noncompact PS sequence, by Theorem 1.6 we know that un → ϕ0 + ki=1 ϕi , where ϕ0 solves the differential equation in (CP) on Ω and the functions ϕi , i = 1, . . . , k, are concentrated solutions on RN with λ = 0. Then Iλ (un ) → Iλ (ϕ0 ) +

k 

I0 (ϕi )

i=1

as n → ∞ and can easily estimate the terms in which the functional is split in the limit. Indeed, by virtue of Proposition 1.2(iii), we have that for every i
1, I0 (ϕi )
N −1 S N/2 . So we can conclude that thesis.

k

i=1 I0 (ϕi )


N −1 S N/2 . Since Iλ (ϕ0 )
0, we have the 

At this stage to conclude the argument regarding the existence of a nontrivial solution to (CP) we only have to show that really infV Iλ < N −1 S N/2 . To this aim we set N
4 and we consider the family of Talenti functions uσ , that is, uσ (x) = σ (N −2)/2 u(σ x), where u(x) = (N (N − 2))(N −2)/4 (1 + |x|2 )(2−N )/2 . We may assume that 0 ∈ Ω and we can choose η ∈ C0∞ (Ω) be a fixed cut-off function such that η = 1 in a neighborhood BR (0) of 0. We set u∗σ = ηuσ . After some computations based on the fact that uσ optimize the Sobolev embedding and so I0 (uσ ) = N −1 S N/2 , we get that, for σ > 0 large ∗ ) < N −1 S N/2 . Indeed, the negative contribution due to the subcritenough, supα∈R σ  Iλ (αu 2 ical term −λ Ω |u| turns out to be less infinitesimal than the positive variation of I0 (ηuσ )

540

S. Solimini

with respect to I0 (uσ ) = N −1 S N/2 , due to the presence of the cut-off function η, when σ is sufficiently large. So we can find α ∈ R such that αu∗σ ∈ V and therefore we have infV Iλ < N −1 S N/2 and infV Iλ > 0 if λ < λ1 . Consequently we get the existence of a nontrivial solution to (CP). 1.6. Existence of a nontrivial solution for the problem on the whole domain Also in this case we shall only give a brief sketch of the proof in a particular case, with the same motivations as in the previous part. Let  

 1 1 2 2 I (u) = |u|p dx |∇u| + a(x)u dx − 2 RN p RN

be the functional associated to the problem (P). Let us assume that lim|x|→+∞ a(x) = a∞ > 0, a(x) < a∞ for every x ∈ RN and let us consider the corresponding functional at infinity  

 1 1 |∇u|2 + a∞ u2 dx − |u|p dx I∞ (u) = 2 RN p RN which is associated to problem

−
u + a∞ u = |u|p−2 u

in RN .

(P∞ )

We introduce the natural constraints V and V∞ related to I and I∞ , respectively. By arguing as in the previous case, since u = 0 is an isolated point for V, we begin by observing that I (u) > 0 ∀u ∈ V. We set c∞ = infu∈V∞ I∞ (u). As we have previously observed, in this case the failure of compactness for a PS sequence can be controlled by scalings diverging by translation and, in order to get the existence result, we have to exclude this case for a constrained minimizing sequence on V. Thus, let (un )n∈N  be a noncompact PS sequence for I ; by Theorem 1.6, we know that I (un ) → I (ϕ0 ) + ki=1 I∞ (ϕi ), where ϕ0 solves the differential equation in (P), and the functions ϕi , i = 1, . . . , k, are solutions to (P∞ ) and the sequences of scalings ρni (ϕi ) are diverging by translation. So, for such a sequence we have I (un ) → c
c∞ and this shows that compactness is achieved for energy levels strictly lower than c∞ . Therefore, to get the existence of a nontrivial solution to (P), it remains to show that infV I (u) < c∞ . To this aim let u∞ be a ground state solution to (P∞ ), fix α ∈ R such that αu∞ ∈ V. Then we have I∞ (u∞ ) = maxα∈R I∞ (αu∞ ) and so I (αu∞ ) < I∞ (αu∞ )  I∞ (u∞ ) = c∞ , from which the condition infV I (u) < c∞ follows.

Multiplicity techniques for problems without compactness

541

2. Approximating problems and compactness of balanced sequences This section is devoted to the proof of the two compactness theorems (Theorems 1.1 and 1.3), which are the most relevant step for proving the multiplicity of solutions (the variational approach through approximating problems discussed in Section 4 presents some difficulties but uses more standard ideas). We shall make use here of some estimates on the solutions which we shall justify roughly, by explaining why we can expect them, but which will be rigorously proved in the next section, which must be therefore considered as an essential part of the present one and which will make the proofs complete. Nevertheless, we prefer to develop the estimates in a separate section at the end, letting the reader already know how they should be and should be used. Let us begin by pointing out that no similar compactness result holds for PS sequences, as stated in the following easy remark. R EMARK 2.1. Problems (CP) and (P) admit noncompact PS sequences. P ROOF. For (CP) we just have to consider the cut-off Talenti function ηuσ considered in the end of Section 1.5 by letting σ = σn → +∞. For (P), when a has limit a∞ at infinity, we just have to fix a solution u¯ to (P∞ ) and take un (x) = u(x ¯ + tn ) with |tn | → +∞.  The impossibility of proving complete compactness results for PS sequences is the reason for which we are working with balanced sequences. The elements of a PS sequence are close to be solutions of the problem while the elements of a balanced sequence are real solutions of a close problem and this makes a big difference: when we deal with the terms of a balanced sequence, we know that we cannot have even a small improvement of some functional of the same type under any local modification of the same order. On the other hand, when some masses are concentrating or escaping to infinity, we shall be able to produce a local modification which improves the value of the functional by respectively perturbing the concentration parameter or the translation vector. From this contradiction we shall be able to deduce the compactness theorems. The variation of the functional under such a local modification will be evaluated by a local Pohozaev inequality and the a priori decay estimates on the terms of a balanced sequence which will be found in the next subsection, carried in such inequalities, will formally produce the contradiction.

2.1. Avoiding concentration In this subsection, we shall test the presence of concentrations which would prevent us to find solutions to (CP) as limits of a balanced concentrating sequence. To this aim, we shall evaluate the infinitesimal variation of the functional corresponding to (SP) under a scaling of a concentrated part of un . Such a variation must be null because we are dealing with a balanced sequence. 2.1.1. Local Pohozaev identity. The property that the variation of the functional under a scaling operation is null or reduced to a boundary term is equivalent to the well-known Pohozaev identity [37] which we must establish in a local form (namely without using

542

S. Solimini

boundary conditions) since it shall be tested on a small concentrated part of the functions un . We fix a general open smooth set B in RN and shall consider, more in general, a semilinear elliptic equation of the form −
u = g(u).

(2.1)

L EMMA 2.1. Let u be a smooth solution to (2.1) on a smooth domain B and let G(u) be a primitive of the function g(u). Then the following equation holds true     1 1 N G(u) + ∗ g(u)u = − |∇u|2 (x · n4) − G(u)(x · n4) 2 B ∂B 2   N + (∇u · x)(∇u · n4) + ∗ (∇u · n4)u. (2.2) 2 ∂B ∂B P ROOF. Multiplying by u and integrating by parts, we get    |∇u|2 = g(u)u + (∇u · n4)u, B

B

(2.3)

∂B

where n4 is the outward normal to ∂B. Multiplying (2.1) for ∇u · x, since



 ∇ · (∇u · x)∇u =
u(∇u · x) + ∇(∇u · x) · ∇u, using the divergence theorem and by integrating by parts, we get  −
u(∇u · x) B

=−



∂B

(∇u · x)(∇u · n4) +



B

 ∇u · ∇ 2 u · x + I · ∇u

 1 |∇u|2 · x + |∇u|2 2 B ∂B B    1 2 −N =− (∇u · x)(∇u · n4) + |∇u|2 (x · n4) + |∇u|2 . 2 ∂B 2 ∂B B =−



(∇u · x)(∇u · n4) +



∇



On the other side, integrating by parts we get    g(u)(∇u · x) = ∇G(u) · x = B

B

∂B

G(u)(x · n4) − N



G(u).

(2.4)

(2.5)

B

Combining (2.18) with (2.5) we obtain    N 2 |∇u| = N G(u) − G(u)(x · n4) 2∗ B B ∂B   1 |∇u|2 (x · n4). (∇u · x)(∇u · n4) + − 2 ∂B ∂B

(2.6)

Multiplicity techniques for problems without compactness

543

Finally, multiplying (2.3) for −N/2∗ and summing (2.6), we have (2.2).



In the present case (i.e., g(u) = |u|p−2 u + λu), (2.2) becomes    N N − ∗ |u|p + λ |u|2 p 2 B B    1 λ = |u|p (x · n4) + |u|2 (x · n4) + (∇u · x)(∇u · n4) p ∂B 2 ∂B ∂B   N 1 |∇u|2 (x · n4) + ∗ (∇u · n4)u. − 2 ∂B 2 ∂B

(2.7)

By a translation, we can move the origin to any point x0 ∈ RN and we can for fixed N N ∗ p get, being p  2 , the positive term ( p − 2∗ ) B |u| , in order to obtain the following “Pohozaev-type” inequality   

 λ

 1 2 p λ |u|  |u| (x − x0 ) · n4 + |u|2 (x − x0 ) · n4 p ∂B 2 ∂B B 

 + ∇u · (x − x0 ) (∇u · n4) ∂B

1 − 2



 N |∇u| (x − x0 ) · n4 + ∗ 2 ∂B 2



∂B

(∇u · n4)u,

(2.8)

which we shall apply to the terms un of a balanced sequence, which enjoys (2.8) for p = pn , on a suitable ball B = Bn . 2.1.2. Decay tools. The choice of the set Bn in (2.8) is a crucial point in order to produce the contradiction. We shall take as Bn a ball around the concentration point xn trying to let Bn contain most of the concentrating mass, making the left-hand side of (2.8) consistent. On the other hand, we must force the right-hand side to be small and, to this aim, we have to take into account two opposite indications: (a) Bn must have a suitably small radius in order to keep the measure of the integration domain ∂Bn small; (b) Bn must have a suitably large radius to keep the points of ∂Bn far away from the concentration in order to make the term u and ∇u which appear in the integrals small. In order to guess a convenient choice of the radius, we can focus on the simple case in which the number k which appears in Definition 1.2 is 1, un = ϕ0 + ρn1 (ϕ1 ) exactly and ϕ1 is a Talenti function (defined at the end of Section 1.5). Since ϕ0 is a smooth function, we easily find the bound ∀n ∈ N, ∀x ∈ Ω:

    un (x)  c 1 +

σn 1 + (σn |x − xn |)2

(N −2)/N

.

(2.9) −1/2

So the values of |un | are of the order of 1 when |x − xn | reaches the order of σn such points we also have the bound ∀n ∈ N, ∀x ∈ Ω:

  un (x)  cσn1/2 .

. In

(2.10)

544

S. Solimini −1/2

So a natural choice is to take σn as the radius of Bn and to bring (2.9) and (2.10) into the right-hand side of (2.8). In the general situation we must take care of several singularities corresponding to unknown functions ϕi and of the difference un − ki=0 ρni (ϕi ) which is infinitesimal only in H 1 . So we shall work on the safe regions Ain which are annuli at −1/2 a distance of the order of σn from the less concentrated mass and avoid the other concentrations. In the next subsection we shall prove the following lemma which essentially gives (2.9) in the most general setting. P ROPOSITION 2.1. Let (un )n∈N be a controlled concentrating sequence. Then there exists a constant C > 0 such that for any n ∈ N and for any x ∈ A2n , un (x)  C. Passing to the smaller annulus A3n we can also give (2.10) in an integral form. P ROPOSITION 2.2. Let (un )n∈N be a controlled concentrating sequence. Then there exists a constant C > 0 such that for any n ∈ N,  (2−N )/2 |∇un |2  Cσn . (2.11) A3n

A simple mean value argument allows us to deduce the following corollary which gives (2.10) in a boundary integral form. C OROLLARY 2.1. Let (un )n∈N be a controlled concentrating sequence. For any n ∈ N  + 2, C  + 3] such that, denoting by Bn the set B(xn , tn σn−1/2 ), there exists tn ∈ [C  (3−N )/2 |∇un |2  Cσn , (2.12) ∂Bn

where C is the constant in the above proposition. R EMARK 2.2. The ball Bn appearing in the previous corollary is not yet, in general, the set on which (2.8) is going to be tested. Indeed, we are not sure that Bn ⊂ Ω. This inclusion is not relevant as far as we work with a controlled sequence, whose terms can be assumed defined on the whole of RN , but must be forced if we want to deal with a balanced sequence in view of applying the Pohozaev inequality (2.8). 2.1.3. Proof of Theorem 1.1. In this subsection we shall use the local Pohozaev identity to prove that concentrations are not possible for balanced sequences in dimension N
7. L EMMA 2.2. If N
7 no concentrating sequence can be balanced. P ROOF. Let a concentrating sequence (un )n∈N be given and assume by contradiction that −1/2 it is balanced. Let us fix n ∈ N, we shall use (2.8) on Bn = B(xn , tn σn ) ∩ Ω, where

Multiplicity techniques for problems without compactness

545

tn is the same as in Corollary 2.1, and we shall split ∂Bn = ∂i Bn ∪ ∂e Bn , where ∂e Bn (empty in the case in which the concentration point xn of the basic rescaled function ϕ is n . When ∂e Bn = ∅, to the aim of applying (2.8), we sufficiently far from ∂Ω) is ∂Ω ∩ B shall take x0 equal to the concentration point xn . Otherwise we shall take x0 out of Ω such −1/2 and that d(x0 , xn )  2tn σn ∀x ∈ ∂e Bn :

n4 · (x − x0 ) < 0,

(2.13)

where n4 is the outward normal to ∂Bn (roughly speaking x0 is the “symmetric” of xn with respect to ∂Ω). We want to show that (2.8) cannot hold true, in contradiction to the assumption that the sequence is balanced. To this aim, we give a lower bound to the lefthand side of (2.8) and a smaller upper bound to the right-hand side. For the first one, we shall restrict the integral on the ball Bn′ = B(xn , σn−1 ), which is contained in Ω for n large. Then we have  

i −1 2 2 −2 un = σn ρn (un )
const, B(xn ,σn−1 )

B(xn ,1)

since, by Remark 1.4, (ρni )−1 (un ) → ϕi = 0 and xn is bounded in RN , we see that the left-hand side of (2.8) has a lower bound of the form Cσn−2 , for a suitable constant C. Passing to the right-hand side, we firstly evaluate the possible contributions of ∂e Bn . On such a set only two of the integrals must be taken into account because we have un = 0 on ∂e Bn ⊂ ∂Ω. For the same reason, ∇un has the direction of n4 and so the whole sum, from (2.13), can be written as  1 |∇un |2 (x − x0 ) · n4  0. 2 ∂e Bn So we can focus our attention to the integrals extended to ∂i Bn . Hence from Proposition 2.1, we get  

 1

 λ |un |2 (x − x0 ) · n4 + |un |p (x − x0 ) · n4 2 ∂i Bn p ∂i Bn 

 −N/2 (x − x0 ) · n4  Cσn , C ∂i Bn

and from Corollary 2.1 and our choice of Bn ,  (2−N )/2 |∇un |2 |x − x0 |  Cσn . ∂i Bn

Finally, from both Proposition 2.1 and Corollary 2.1, by the Hölder inequality, 

∂i Bn

(∇un · n4)un 



∂i Bn

|∇un |2

1/2 

∂i Bn

|un |2

1/2

(2−N )/2

 Cσn

.

546

S. Solimini

Combining these estimates, we see that the right-hand side of (2.8) is therefore bounded (2−N )/2 . So (2.8) requires by Cσn (2−N )/2

λσn−2  Cσn

,

which when N > 6, since σn → ∞, is clearly false for n large.

(2.14) 

Theorem 1.1 is an immediate consequence (essentially a restating which does not use the terminology introduced in this chapter) of the above lemma. P ROOF OF T HEOREM 1.1. Let us suppose, by contradiction, that there exists a bounded balanced sequence (un )n∈N such that   sup sup un (x) = +∞. n∈N x∈Ω

A standard regularity argument [10] shows that un cannot be compact in H 1 , so by Lemma 1.1 it has a balanced concentrating subsequence and this is excluded by Lemma 2.2.  2.2. Avoiding escaping masses

We shall now take into exam the case of a balanced sequence related to problem (P) and we shall assume by contradiction that the sequence is broken, according to Definition 1.3. This means, roughly speaking, that there are some masses ϕi which are escaping to infinity. So we are going to study the variation of the functional under a small translation of one of such masses which brings it back to the origin. 2.2.1. Local Pohozaev identity for translations. The variation of the functional under the translation of a solution is evaluated by a Pohozaev-type formula. Since we only want to translate a part of the function, corresponding to one of the escaping masses, we must prove such a formula in a local version, namely without assuming boundary conditions, as stated in the next lemma. We fix a general open smooth set B in RN and shall consider, more in general, a semilinear elliptic equation of the form −
u = g(x, u).

(2.15)

L EMMA 2.3. Let u be a smooth solution to (2.15) on a smooth domain B and let G(x, s) be a primitive with respect to s of the function g(x, s). Then the following equation holds true  − ∇x G(x, u) · t4 B

=



∂B







 1 2 4 |∇u| − G(x, u) ν · t − (∇u · ν) ∇u · t4 . 2 ∂B

(2.16)

Multiplicity techniques for problems without compactness

547

P ROOF. We have 

  −
u + g(x, u)u ∇u · t4 dx = 0.

(2.17)

B

Now integrating by parts, we obtain 

B

 −
u ∇u · t4 dx =



B

 ∇u · ∇ ∇u · t4 dx −



∂B

 (∇u · ν) ∇u · t4 dσ .

Then, taking into account that t4 does not depend on x, again using divergence theorem, we get 

B

 ∇u · ∇ ∇u · t4 dx =



=

B

1 2

1 = 2

 ∇u · ∇ 2 u · t4 dx



B



 ∇|∇u|2 · t4 dx



|∇u|2 t4 · ν dσ

∂B

and then 

B

 −
u ∇u · t4 dx

1 = 2



∂B



|∇u|2 t4 · ν dσ −



∂B

 (∇u · ν) ∇u · t4 dσ .

(2.18)

Analogously we deduce 



G(x, u)ν · t4 =



∇G(x, u) · t4 =



g(x, u)∇u · t4 =



G(x, u)ν · t4 −

∂B

B

B

∇x G(x, u) · t4 +



B

g(x, u)∇u · t4.

So

B

∂B



B

∇x G(x, u) · t4.

Therefore 1 2



∂B

=





|∇u|2 t4 · ν −

∂B



G(x, u)ν · t4 −

∂B



B

 (∇u · ν) ∇u · t4 ∇x G(x, u) · t4.



548

S. Solimini

In the case of problem (P) we can take G(x, s) = p1 |s|p −

a(x) 2 2 s .

C OROLLARY 2.2. Let a(x) satisfy (a1 ) and u be a solution to (P). Then the following identity 1 2



 1 u ∇a(x) · t4 dx = 2 B 2



−

∂B



  |∇u|2 + a(x)u2 ν · t4

 1 (∇u · ν) ∇u · t4 − p ∂B



∂B

 |u|p ν · t4 ,

(2.19)

where ν is the outward normal to ∂B, holds. 2.2.2. Drift estimates tools. As for the case of (CP), we must now choose a convenient set B on which (1.10) leads to a contradiction. Again the contradiction will follow from the fact that the boundary integrals are too small with respect to the volume integral and this analysis is based on suitable decay estimates. Solutions to (P) have an exponential decay at infinity. One can guess that therefore we should find an uniform exponential decay on the terms of a balanced sequence of functions if we keep far away from the escaping masses. This means that the bound we are going to find is not of the type of e−α|x| but of e−ασn (x) , where the function σn defined by   σn (x) = inf x − tni , 0ik

x ∈ RN ,

(2.20)

will be called drift distance function and measures how much x is escaping from all the masses in which un gets broken. Note that tn0 = 0 so σn (x)  |x| for all n ∈ N. Indeed in Section 3 we shall prove the following exponential decay result. P ROPOSITION 2.3. Let a(x) satisfy (a1 ) and (a2 ). Let (un )n∈N be a broken controlled √ sequence bounded in H 1 (RN ). Then for any constant α ∈ (0, a∞ ), there exists a constant cα > 0 such that for n large enough, un (x)  cα e−ασn (x)

∀x ∈ Rn .

(2.21)

The above estimate suggests to apply (2.19), where u = un is a term of a balanced sequence of functions, to a set B = Dn whose boundary is far away from all the masses, as it happens with the cones Cn and the other regions introduced in the end of Section 1.3. We must take into account that un solves the problem on Bρn , so we have to take the trace of such cones on Bρn . As in the case of the critical growth problem, the choice of Bn will follow from a gradient estimate in an integral form. Indeed, in Section 3 we shall also prove the following estimate.

Multiplicity techniques for problems without compactness

549

P ROPOSITION 2.4. Let a(x) satisfy (a1 ) and (a2 ). Let (un )n∈N be a broken controlled sequence bounded in H 1 (RN ). Then there exist constants α∗ > 0 and c∗ > 0 such that for all n ∈ N, 

S1,n

|∇un |2 dx  c∗ e−α∗ |tn | ,

(2.22)

where S1,n is as defined in (1.13). Also in this case a mean value argument allows to pass to a boundary integral. P ROPOSITION 2.5. Let a(x) and (un )n∈N be as in Proposition 2.3. Then there exist constants α ∗ > 0, c∗ > 0 and a sequence (sn )n∈N , sn ∈ (− 21 , 12 ) such that for all n ∈ N, 

∂ Csn ,n

|∇un |2 dx  c∗ e−α

∗ |t

n|

,

(2.23)

where, for all n, Csn ,n is as defined in (1.12). Then, it makes sense setting Dn = Cn ∩ Bρn (0),

where Cn denotes the cone Csn ,n . We remark that, for large n, the vertex t2n ∈ Dn ; indeed, even if ρn < |tn |, for all n, |tn | − ρn  C for some constant C, otherwise un (· − tn ) ⇀ 0 n→+∞

contradicting the choice of tn . Moreover, we remark that ∂Dn consists of an “internal part” (∂Dn )i = ∂ Cn ∩ Bρn (0)

and an “external” one

(∂Dn )e = Cn ∩ ∂Bρn (0).

We finally point out that by using (2.21) we easily get the following integral estimate, whose detailed proof is in Section 3. P ROPOSITION 2.6. Let a(x) satisfy (a1 ) and (a2 ). Let (un )n∈N be a broken controlled sequence bounded in H 1 (RN ) and Sn be as in (1.14). Then, for all p
2, there exist constants α˜ > 0 and c˜ > 0 such that for n large enough, 

Sn

˜ n| (un )p dx  ce ˜ −α|t .

(2.24)

550

S. Solimini

2.2.3. Proof of Theorem 1.3. We are now ready to combine the local Pohozaev formula (2.19) with the exponential decay estimates in the previous propositions to the aim of proving Theorem 1.3. Let us begin with a lemma which is the only step in which we use (a4 ). L EMMA 2.4. Let a(x) satisfy (a1 ) and (a4 ). Let (un )n∈N be a noncompact balanced sequence. Then, for large n, the inequality 

 

∂a 1 (x)u2n dx ∇a(x) · t4n u2n dx
2 Dn ∂ x4 Dn

(2.25)

holds. P ROOF. Denoting by (t4n )τx the component of t4n lying in the space orthogonal to x4 and containing x, using (a4 ) we get, for large n,   

 

∇a(x) · t4n = ∇a(x) · x4 t4n · x4 + ∇τx a(x) · t4n τ x

 ∂a ∂a   
(x) t4n · x4 − c¯ (x) t4n τ  x ∂ x4 ∂ x4  

 &  ∂a % (x) t4n · x4 − c¯ t4n τ  . = x ∂ x4

In order to evaluate [(t4n · x4) − c|( ¯ t4n )τx |], let us first suppose x ∈ B2Rn (tn ), so that |x − tn | < 2Rn < γˆ |tn |, with γˆ as in (1.11), then we have 

t4n · x4 =



tn t n + x − tn · |tn | |x|






|tn | − |x − tn | |tn | − |x − tn | 1 − γˆ

|x| |tn | + |x − tn | 1 + γˆ (2.26)

and since tn − x x + , |tn | |tn |    |tn − x|  t4n   < γˆ . τx |x| t4n =

(2.27)

On the other hand, we can assert that, by homothety, (2.26) and (2.27) are also true for all x belonging to the cone K having as vertex the origin and generated by B2Rn (tn ). Then, in particular, (2.26) and (2.27) are true for all x ∈ Dn , being Dn ⊂ Cn ⊂ K. γˆ 1−γˆ > 21 and 1+ − 4c¯γˆ > 0. Thus (2.25) follows because we have, by the choice of γˆ , 43 1− 1+γˆ γˆ  We can combine the previous lemma with Corollary 2.2 obtaining the following result.

Multiplicity techniques for problems without compactness

L EMMA 2.5. Let a(x) and (un )n∈N be as in Lemma 2.4. Then the inequality  1 ∂a (x)u2n (x) 4 Dn ∂ x4 

  1  |∇un |2 + a(x)u2n νn · t4n 2 (∂ Dn )i  



 1 − |un |p νn · t4n (∇un · νn ) ∇un · t4n − p (∂ Dn )i (∂ Dn )i

551

(2.28)

holds. P ROOF. Combining (2.19) and (2.25) we obtain  

  1 ∂a 2 1 un  |∇un |2 + a(x)u2n νn · t4n 4 Dn ∂ x4 2 ∂ Dn  

 1

 − (∇un · νn ) ∇un · t4n − |un |p νn · t4n . p ∂ Dn ∂ Dn

(2.29)

Now, for all n, un solves (PBρn (0) ), un = 0 on ∂Bρn (0) ⊃ (∂Dn )e , so ∇un and νn have the same direction, moreover, on (∂Dn )e it is (νn · t4n )
0, thus we deduce  



 2 4 |un |p νn · t4n (2.30) a(x)un νn · tn = 0 = (∂ Dn )e

(∂ Dn )e

and 1 2



(∂ Dn )e

1 = 2



1 =− 2

 |∇un |2 νn · t4n −



(∂ Dn )e

 |∇un | νn · t4n − 2

(∂ Dn )e



(∂ Dn )e

 (∇un · νn ) ∇un · t4n

 1 (∇un · θ ∇un ) νn · t4n θ (∂ Dn )e



 |∇un |2 νn · t4n  0.

Hence (2.28) follows inserting (2.30) and (2.31) in (2.29).

(2.31) 

Taking the decay estimate in (2.28) we can finally deduce the compactness of a balanced sequence. L EMMA 2.6. Let a(x) satisfy (a1 )–(a4 ) and (un )n∈N be a bounded balanced sequence. Then (un )n∈N is relatively compact. P ROOF. We argue by contradiction and we assume that (un )n∈N is not compact. Then, by Proposition 1.1, up to a subsequence, it is broken in Lp and, by Lemma 2.5, the inequality (2.28) must be true.

552

S. Solimini

Let us consider, for n large, the right-hand side of (2.28). First of all, let us observe that, by (a2 ), a(x)
0 for all x ∈ (∂Dn )i so, taking into account that (νn · t4n )  0 on (∂Dn )i , we have 

(∂ Dn )i

  |∇un |2 + a(x)u2n νn · t4n  0.

(2.32)

Moreover, by using Proposition 2.5, we deduce −



(∂ Dn )i







(∇un · νn ) ∇un · t4n dσ 

2

(∂ Dn )i

|∇un | dσ 

∂ Cn

|∇un |2 dσ  c∗ e−α

∗ |t

n|

.

(2.33)

Let us now show that there exist constants α ′ > 0 and c′ > 0, independent on n, so that −



(∂ Dn )i

 |un |p νn · t4n dσ 



(∂ Dn )i

′

|un |p dσ  c′ e−α |tn | .

(2.34)

Since (∂Dn )i ⊂ ∂ Cn and, for large n, ∂ Cn ⊂ Sn , using Proposition 2.3 we infer −



(∂ Dn )i





∂ Cn

|un |p dσ

|un |p dσ  cα



∂ Cn

e−ασn (x)p dσ  cα

√ α ∈ (0, a∞ ), cα > 0. Setting, for h
1 and i = 0, 1, . . . , k, Ah,i


  i h rn h−1 rn   < x − tn < 2 = x ∈ ∂ Cn : 2 2 2

k  

 i=1 ∂ Cn

i

e−αp|x−tn | dσ ,

(2.35)

(2.36)

and denoting by |Ah,i | the (N − 1)-dimensional (Hausdorff ) measure of Ah,i , we have for i = 0, 1, . . . , k, (N −1 ' h rn , |Ah,i |  C 2 2

C ∈ R,

(2.37)

because it is not difficult to understand that, ∀h, |Ah,i | can be estimated by the surface of the cylinder having height and basis diameter measure equal to r2n 2h .

Multiplicity techniques for problems without compactness

553

Thus, in view of (2.36) and (2.37), we deduce 

∂ Cn

e

−αp|x−tni |

dσ 

∞  

h−1 r

e−αp2

n /2

dσ

h=1 Ah,i

C

∞ 

h−1 r

e−αp2

n /2

h=1

' ( rn N −1 2h 2

(2.38)

hence, inserting (2.38) in (2.35), we obtain as desired, 

(∂ Dn )i

|un |p dσ  cα′ krnN −1 e−αprn /2

∞ 

h

′

e−αp2 2h(N −1)  c′ e−α |tn | .

(2.39)

h=0

On the other hand, denoting by ρ˜n = max{ρn , |tn |} and by n = Cn ∩ Bρ˜ (0), D n

we have, for large n,

    ∂a ∂a ∂a 2 2  (x)un dx
inf (x) (x) u2n dx, un dx
C inf   ∂ x 4 ∂ x 4 ∂ x 4 D n Dn Dn Dn Dn



(2.40)

 > 0 constant, because, as remarked at the beginning of the section, (||tn | − ρn |)n∈N is C bounded from above. Moreover, in view of Proposition 1.1 and of the choice of tn , we infer  u2n dx
λ > 0, λ = const. (2.41) lim inf n→+∞ D n

Then, combining (2.28) with (2.32)–(2.34), (2.40) and (2.41), we obtain  1  ∂a c′ ∗ ′ ¯ n| λC inf (x)  c∗ e−α |tn | + e−α |tn |  ce ¯ −α|t ,  4 ∂ x 4 p Dn

α¯ = min(α ∗ , α ′ ), and this is impossible by (a3 ).



P ROOF OF T HEOREM 1.3. If the statement is not true, we can extract from U a balanced sequence which is not precompact and therefore, by Proposition 1.1, has a broken subsequence. Then we get in contradiction to the previous lemma.  3. Decay estimates This section completes the previous one by proving the estimates stated in Sections 2.1.2 and 2.2.2.

554

S. Solimini

3.1. Decay estimates near concentration points We start by considering a balanced sequence related to problem (CP) and establishing the estimates in Section 2.1.2. We recall that we can always substitute the terms un with their absolute value, extended by 0 on all of RN , passing to the weaker assumption that the sequence is controlled but getting free from caring about the sign of the function or the shape of Ω. 3.1.1. Integral estimates for controlled concentrating sequences. So we shall work with a controlled concentrating sequence (un )n∈N . The boundedness of the sequence in H01 , and ∗ so in L2 , cannot hold in Lp for p > 2∗ , because of the presence of concentrations. On the other hand, such concentrations are small in Lp for p < 2∗ , so that, modulo an infinitesimal term, un can be split in a part which keeps bounded in Lp for large p and in a part which is infinitesimal in Lp for small p. In order to guess what kind of estimate we are going to find, we can assume that un = ϕ0 + ρn1 (ϕ1 ) exactly and that ϕ1 is a Talenti function, as in the beginning of Section 2.1.2. In such a case, ϕ0 ∈ Lp for every p, while ϕ1 ∈ Lp only for p > 2∗ /2 = N/(N − 2). If p1 > 2∗ , then ϕ0 p1  const and if 2∗ /2 < p2 < 2∗ , then   1 ρ (ϕ1 ) n

p2

= σn1

N/2∗ −N/p2

ϕ1 p2 = σn1

N/2∗ −N/p2

.

If one has several concentrating masses ϕi , then  k      i ρn (ϕi )    i=1

k    i ρ (ϕi )



n

p2

i=1

p2

k 



σni

N/2∗ −N/p2

i=1

ϕ1 p2

N/2∗ −N/p2

 const · σn

.

(3.1)

So we are lead to introduce the following definition. D EFINITION 3.1. Let p1 , p2 ∈ ]2, +∞[ be real numbers such that p2 < 2∗ < p1 , α > 0 and σ > 0. We consider an inequalities system


u1 p1  α,

∗ −N/p

u2 p2  ασ N/2

2

,

(3.2)

which will let us introduce a norm depending on p1 , p2 and σ , by setting  up1 ,p2 ,σ = inf α > 0 | ∃u1 , u2 such that (3.2) is satisfied and |u|  u1 + u2 .

The above norm will be briefly denoted by uσ when p1 and p2 can be supposed to be given.

Multiplicity techniques for problems without compactness

555

R EMARK 3.1. Let p1 , p2 ∈ ]2, +∞[ real numbers such that p2 < 2∗ < p1 and σ > 0, then, by definition, for any function u, we get uσ  up1 ,

∗

uσ  up2 σ N/p2 −N/2 .

We can easily see from (3.1) that for every p1 , p2 such that   k      i ρn (ϕi ) ϕ0 +   i=1

2∗ 2

< p2 < 2 ∗ < p1 ,

 const. p1 ,p2 ,σn

The main goal  of this section is to show that the same bound holds for un , which differs from ϕ0 + ki=1 ρni (ϕi ) by an infinitesimal term in H01 which is not a priori even bounded in the norm  · p1 ,p2 ,σn . Thanks to the assumption that (un )n∈N is also controlled, we shall be able to prove the following Brezis–Kato-type regularity result (see [10], Theorem 2.3): P ROPOSITION 3.1. Let (un )n∈N be a controlled concentrating sequence, then for any ∗ p1 , p2 ∈ ] 22 , +∞[, p2 < 2∗ < p1 , there exists a constant C(p1 , p2 ) depending on the sequence and on the exponents p1 and p2 , such that for any n ∈ N, un σn  C. To this aim, we shall state three preliminary lemmas: a continuity lemma, a bootstrap lemma and the relative initialization lemma. L EMMA 3.1. Let u and v ∈ H 1 (RN ) and a ∈ LN/2 (RN ) be three positive functions such that −
u  a(x)v. Then for each p1 , p2 ∈ ]2, +∞[, there exists a constant C(N, p1 , p2 ), depending on the dimension N and on the exponents p1 and p2 , such that for any σ > 0, uσ  C(N, p1 , p2 )aN/2 vσ . P ROOF. Let u, v be as in the statement of the lemma and let fix σ > 0 and ε > 0. Let v  v1 + v2 such that v1 and v2 satisfy (3.2) for α = vp1 ,p2 ,σ + ε. Let us consider, for i = 1, 2, the solution ui ∈ H 1 (RN ) to −
ui = avi . Then ui pi  C(N, pi )aN/2 vi pi and, being −
u1 −
u2 = av1 + av2
av
−
u, by the maximum principle, we have u  u1 + u2 . Since the functions ui satisfy (3.2) with α = C(N, p1 , p2 )aN/2 × (vσ + ε), with C(N, p1 , p2 ) = max(C(N, p1 ), C(N, p2 )), by the arbitrariness of ε we get the thesis.  The bootstrap argument relies in the use of the following lemma.

556

S. Solimini

N +2 N L EMMA 3.2. Let p1 , p2 ∈ ] N −2 , 2 i = 1, 2, by

N +2 N −2 [

such that p2 < 2∗ < p1 and let qi be defined, for

N +2 1 2 1 = − . qi N − 2 pi N

(3.3)

If u and v are two positive functions whose support is contained in a bounded set Ω and such that ∗ −1

−
u  v 2

+ A,

then there exists a constant C(N, p1 , p2 , Ω) such that for any σ > 0,

(N +2)/(N −2)  +1 . uq1 ,q2 ,σ  C(N, p1 , p2 , Ω) vp1 ,p2 ,σ

(3.4)

P ROOF. By proceeding as in the previous lemma, we consider v = v1 + v2 where the functions vi satisfy (3.2) for α = vp1 ,p2 ,σ + ε and ε is a real strictly positive number arbitrarily small. Let u1 and u2 be two functions in H01 (Ω) such that (N +2)/(N −2)

−
u1 = 24/(N −2) v1

(N +2)/(N −2)

−
u2 = 24/(N −2) v2

+ A, .

Since −
u  v (N +2)/(N −2) + A

(N +2)/(N −2)

 2(N +2)/(N −2)−1 v1

(N +2)/(N −2)

+ A + 2(N +2)/(N −2)−1 v2

= −
u1 −
u2 , by the maximum principle, u  u1 + u2 follows. Hence, we have to estimate u1 q1 N +2 N N +2 and u2 q2 . We have, using (3.3) and being N −2 < pi < 2 N −2 ,   (N +2)/(N −2) u1 q1  C(N, p1 )v1 + ALp1 (N−2)/(N+2)

 (N +2)/(N−2)  C(N, p1 ) v1 p1 + A|Ω|1/p1 (N +2)/(N −2)

(N +2)/(N −2)  +1 .  C(N, p1 , Ω) vp1 ,p2 ,σ + ε

Analogously, if we use the equality N N − = 2∗ q2



N N N +2 , − 2 ∗ p2 N − 2

Multiplicity techniques for problems without compactness

557

we get (N +2)/(N −2)

u2 q2  C(N, p2 ) v2 p2 %  &(N +2)/(N −2) ∗  C(N, p2 ) vp1 ,p2 ,σ + ε σ N/2 −N/p2

(N +2)/(N −2) (N/2∗ −N/p )(N +2)/(N −2) 2 = C(N, p2 ) vp1 ,p2 ,σ + ε σ

(N +2)/(N −2) N/2∗ −N/q 2. = C(N, p2 ) vp1 ,p2 ,σ + ε σ

So u1 and u2 solve (3.2) for C = C(N, p1 , p2 , Ω)((vp1 ,p2 ,σ + ε)(N +2)/(N −2) + 1); this concludes the proof by the arbitrary choice of ε.  Now we need to initialize the exponents through the following lemma. L EMMA 3.3. Let (un )n∈N be a controlled concentrating sequence then there exists a con∗ stant C and exponents p1 , p2 ∈ ] 22 , +∞[, p2 < 2∗ < p1 , such that for any n ∈ N, un σn  C.

(3.5)

P ROOF. This proof will follow a Brezis–Kato-type argument (see [10]) in order to get free from an infinitesimal term which is the only real obstacle to our estimates, as explained in the beginning of this section. For any n ∈ N, we can consider using a homogeneous notation, un = u0n + u1n + u2n , where • u1n stands for the weak limit ϕ0 ,  • u2n stands for the sum of rescaled functions ϕi , u2n = ki=1 ρni (ϕi ), • u0n = un − u2n − ϕ0 is, by definition of concentrating sequence, an infinitesimal term ∗ in L2 norm. We shall overcome the difficulty due to the presence of u0n by taking advantage of the assumption that we are dealing with a controlled concentrating sequence. Let u be one of the terms un , ui = uin and ai = max(1, 3(6−N )/(N−2) )ui 4/(N −2) for i = 1, 2, 3, and σ = σn . The infinitesimal character of u0n shall allow us to consider a0 as small as we want in the LN/2 norm ((3.5) is easily checked on a finite number of terms, see [10] and [25]). Being ∗ −2

a = u2



4/(N −2) 4/(N −2)  + u2  max 1, 3(6−N )/(N−2) |u0 |4/(N −2) + u1 ,

we can consider u as a solution to −
u  (a0 + a1 + a2 )u + A, so by the monotonicity of the Green operator G (G : H −1 (Ω) → H01 (Ω) denotes the inverse operator of −
) we have u  G(a0 u) + G(a1 u + A) + G(a2 u).

(3.6) ∗

Since Ω is a bounded set and a1 ∈ L∞ , we get that G(a1 u+A) is bounded in W 2,2 ֒→ Lp1 for any p1 such that N −6 1 2 1
− = p1 2 ∗ N 2N

558

S. Solimini

and so (see Remark 3.1)     G(a1 u + A)  G(a1 u + A)  C. σ p

(3.7)

1

Now let 2∗′ < p2 < 2∗ be given. We consider the index r such that 1 2 1 1 = + ∗− , p2 r 2 N from p2 > 2∗′ we get r > N4 . The decay speed of the solution ϕ = ϕi (see [25]) allows us to say that a2 ∈ Lr and, if we want to estimate the Lr norm of a2 , we just have to take into account the less concentrated term, namely ρn (ϕ), as follows from r < N2 , which is in turn a consequence of p2 < 2∗ . By easy computations we have a2 Lr  Cσ 2−N/r , which, taking into account that 2 − N/r = N/2∗ − N/p2 , implies   G(a2 u)

p2

∗ −N/p

 Ca2 Lr uL2∗  Cσ N/2

2

,

(3.8)

therefore, from Remark 3.1,     G(a2 u)  σ N/p2 −N/2∗ G(a2 u)  C. p σ 2

(3.9)

Now we point out that with the above choice for p1 and p2 we get   G(a0 u)  1 uσ . σ 2

(3.10)

Indeed, by Lemma 3.1, we get

  G(a0 u)  Ca0 N/2 uσ  1 uσ , σ 2

(3.11)

under a suitable choice of the bound on the norm of a0 . So by (3.6), (3.10) and the triangular inequality, we finally obtain     uσ  2G(a1 u + A)σ + 2G(a2 u)σ ,

which, combined with (3.7) and (3.9), gives the thesis.



P ROOF OF P ROPOSITION 3.1. Let (un )n∈N be a controlled concentrating sequence. By applying the initialization Lemma 5.2, we can find a constant C > 0 and two exponents, N +2 N N +2 ∗ p1 and p2 ∈ ] N −2 , 2 N −2 [, p2 < 2 < p1 , such that (3.5) holds. Using the bootstrap Lemma 3.2 we can repeatedly enlarge the interval ]p2 , p1 [ to ]q2 , q1 [, where the expo-

Multiplicity techniques for problems without compactness

559

nents qi are given by (3.3), obtaining (3.4). This procedure allows us to manage, in a finite ∗ number of steps, every exponent p1 , p2 ∈ ] 22 , +∞[.  3.1.2. Local uniform bounds on controlled concentrating sequences. We are now going to establish the local uniform bound on the terms of a controlled concentrating sequence on the safe regions A2n stated in Proposition 2.1, whose proof is the main goal of this section. The proof is a simple variant of the argument used in [41] and in [25] and shall require some preliminary steps. We begin by establishing a weaker estimate. P ROPOSITION 3.2. Let (un )n∈N a controlled concentrating sequence. Then there exists a constant C > 0 such that, for any n ∈ N and for any x ∈ A1n , (N −2)/4

un (x)  Cσn

.

P ROOF. We shall proceed by contradiction: let (yn )n∈N be a sequence such that yn ∈ A1n for any n ∈ N and (2−N )/4

lim un (yn )σn

n→+∞

= +∞,

(3.12)

and let us scale the functions un in such a way to carry the point yn in the origin and normalize the value of the functions. The required scaling sends un in u˜ n defined as N/2∗

u˜ n (x) = ρn

un (ρn x + yn ),

where

2/(2−N ) −2∗ /N ρn = un (yn ) = un (yn ) ,

so that u˜ n (0) = 1. Note that, using (3.12), we have lim

n→+∞

ρn −1/2 σn

= 0.

(3.13)

Therefore, since yn ∈ A1n , there is no concentration point which approximates yn at a dis−1/2 tance less than or equal to σn and so of the order of ρn , we can deduce that u˜ n ⇀ u˜ = 0. The contradiction will be archived by showing that we can choose the points yn in such a −1/2 way to have u˜ = 0. This shall possibly force us to work on a (εσn )-neighborhood of A1n , but this change will obviously not make any relevant difference in the above argument. The choice will consist in forcing the property u˜ n (y)  2

 = 2u˜ n (0)

∀y ∈ Bρ (0),

(3.14)

560

S. Solimini

for some given ρ > 0. Then by using that u˜ n still satisfies (EI) and by estimating the variation of the mean value of un , we have for 0 < r  ρ,    r 1 − u˜ n = u˜ n (0) +
u ˜ n dt N −1 ∂Br Bt 0 N bN t  r 

2∗ −1  1 1
1−C 2 + A dt = 1 − Cr 2
, N −1 2 Bt 0 t where bN stands for the (N − 1)-dimensional measure of the unit sphere in RN , provided we choose r conveniently small. So the weak limit u˜ cannot be zero. Therefore we only have to prove (3.14). To this aim, let us fix ρ > 0 and assume that, for a given n ∈ N, yn does not satisfy (3.14). Then we must fire yn and look for a better point to hire for the same job. Since (3.14) is false, we can find zn ∈ Bρ (0) such that (N −2)/2

u˜ n (zn ) = ρn

un (ρn zn + yn )
2.

(3.15)

The first candidate to replace yn is yn(1) = ρn zn + yn which leads us to replace ρn by &2/(2−N ) % ρn(1) = un yn(1)  22/(2−N ) ρn .

(3.16)

(1)

We can be sure that yn is at least as good as yn to let (3.12) hold since (3.15) implies that

 un yn(1)
2un (yn ).

(3.17)

Moreover, being zn ∈ Bρ (0), we get

 (1)  y − yn  = |zn ρn |  ρρn .

(3.18)

n

(1)

(1)

We can define u˜ n as before by substituting yn and ρn with yn and ρn , respectively. If this new u˜ n satisfies (3.14) we do not have to look for other choices. Otherwise, we repeat (2) the same argument and we choose a second candidate yn by arguing in the same way. For any fixed n ∈ N, we proceed recursively finding a sequence yn(1) , yn(2) , . . . , yn(k) , . . . as far as we do not find a successful choice, which lets us claim (3.14). We can easily see that this process cannot go on indefinitely. Indeed (3.16) becomes in the general case, for i > 0, ρn(i+1)  22/(2−N ) ρn(i) and (3.18),   (i+1) y − yn(i)   ρρn(i) . n

Multiplicity techniques for problems without compactness

561 (i)

Then one easily sees, by taking the sum of a geometric sequence, that yn converges to a (∞) (i) point yn as i → +∞ but, by construction, we have un (yn ) → +∞, in contradiction to the smoothness of un . Finally, for every i > 0, we have +∞    (i) −1/2 y − yn   ρρn 22/(2−N )j < εσn n j =0

(i)

−1/2

for n large. So all the points yn are in the (εσn used to replace yn .

)-neighborhood of A1n and so can be 

P ROPOSITION 3.3. Let (un )n∈N be a controlled concentrating sequence, then there exists  n−1/2 , (C  + 5)σn−1/2 ], a constant C > 0 such that, for any n ∈ N and for any r ∈ [Cσ  − un  C. ∂Br (xn )

∗

P ROOF. By continuity, being (un )n∈N bounded in L2 ⊂ L1 , we can suppose  un  C B1 (xn )

with a constant C independent from n. So, for any n ∈ N, there exist rn ∈ [ 21 , 1], such that  −

∂Brn (xn )

un = C.

N +2 N +2 We are going to use Proposition 3.1 for p1 = N N −2 and p2 = N −2 , so, for any n ∈ N, we choose u1 = u1,n and u2 = u2,n such that (3.2) is satisfied for σ = σn and with a constant α that does not depend on n. Estimating the spherical mean variation from rn to r and taking  + 5)σn−1/2 < 1/2, i.e., r < rn for n large, we find into account that (C    rn   r 1 d − − un dt = C + −
un dt un = C + N bN t N −1 Bt (xn ) r rn dt ∂Bt (xn ) ∂Br (xn )  1 

2∗ −1  1 C+ + A dt un N −1 −1/2 N bN t  n Cσ Bt (xn )  1  1 (N +2)/(N −2) 4/(N −2) C+ u dt 2 N bN t N −1 Bt (xn ) 1,n  n−1/2 Cσ  1   1 A 1 (N +2)/(N −2) 4/(N −2) + t dt 2 u dt + N 0 N bN t N −1 Bt (xn ) 2,n  n−1/2 Cσ

=C+

24/(N −2) A , (A1 + A2 ) + N bN 2N

562

S. Solimini

where, for i = 1, 2, Ai =



1

1

 n−1/2 t N −1 Cσ



(N +2)/(N −2)

Bt (xn )

ui,n

dt.

Being u1,n ∈ LN (N +2)/(N −2) , by the Hölder inequality, we get A1  C



1 0

1 N 1−1/N (N +2)/(N −2) t u1,n LN(N+2)/(N−2) dt  Cα  C.

t N −1

(N +2)/(N −2)

On the other side, being u2,n ∈ L(N +2)/(N −2) , i.e., u2,n A2 



1

 n−1/2 Cσ

∈ L1 we have

1 % (N/2∗ −N (N −2)/(N +2)) &(N +2)/(N −2) ασn dt

t N −1

(2−N )/2

= α (N +2)/(N −2) σn



1

 n−1/2 Cσ

1

t N −1

dt  C, 

and this concludes the proof. From Proposition 3.3 we see, by integrating with respect to r, that  − un  C.

(3.19)

A1n

Since, ∀x ∈ A2n : Bσ −1/2 (x) ⊂ A1n and the measure of the two sets are of the same order, n from (3.19) we deduce that  ∀x ∈ A2n : − un  C. (3.20) B

−1/2 (x) σn

Since  un (x) = lim −

ρ→0 Bρ (x)

un ,

 Proposition 2.1 follows from (3.20) if we estimate the variation of −Bρ (x) un for 0  −1/2

ρ  σn

.

P ROOF OF P ROPOSITION 2.1. Let us fix an index n ∈ N and a point x ∈ A2n . If un (x)   2 −B −1/2 (x) un , by (3.20) we have done. Otherwise, setting for any ρ > 0, σn

 m(ρ) = −

∂Bρ (x)

un

and m(0) = un (x),

Multiplicity techniques for problems without compactness

563

we deduce that −1/2

∃ρ¯  σn

1 1 such that m(ρ) ¯  m(0) = un (x). 2 2

Then we take ρ1 and ρ2 ∈ [0, ρ] ¯ such that m(ρ) attains its maximum in ρ1 , and ρ2 is the least value of ρ
ρ1 such that m(ρ)  12 m(ρ1 ). Being un solution to (EI), and Bρ2 (x) ⊂ A1n , we have on such a set, by Proposition 3.2, 4/(N −2)  Cσn . So we find, for n sufficiently large, un ρ1 

   ρ2 d 1 − −
un dρ un dρ = N −1 dρ ∂Bρ (x) Bρ (x) ρ1 N bN ρ ρ2   ρ2

4/(N −2)  1 un un + A dρ  N −1 Bρ (x) ρ1 N bN ρ 3  ρ2 4 1 1 N 4/(N−2) + Ab dρ  sup u u ρ n n N N bN ρ1 ρ N −1 Bρ (x) Bρ (x)    ρ2 1 N C σn un + Aρ dρ N −1 Bρ (x) ρ1 ρ   ρ2 

ρ dρ  Cm(ρ1 )σn ρ22 − ρ12 ,  C m(ρ1 )σn + A

1 m(ρ1 ) = 2



ρ1

−1/2

therefore (ρ22 − ρ12 ) > Cσn−1 and so ρ2 − ρ1 > Cσn . Denoting by A the annulus cen−N/2 tered in x of radii ρ1 and ρ2 we have that measure of A is of the order of σn , i.e., of the same order of A1n and so as in (3.20) we have  − un  C. A

On the other hand,  1 − un
m(ρ2 ) = m(ρ1 ) 2 A and so un (x) = m(0)  m(ρ1 )  C.



3.1.3. Gradient estimates. In this subsection we shall prove the integral bound for the derivatives of every term un of a controlled concentrating sequence in its safe regions A3n stated in Proposition 2.2. One can easily guess that, since un and
un are uniformly −1/2 bounded on A2n and the width of A2n is of the order of σn , ∇un can be expected to

564

S. Solimini 1/2

be of the order of σn as stated in an integral form in (2.11). Such an estimate can be easily proved in a rigorous way by a Caccioppoli-type inequality. P ROOF OF P ROPOSITION 2.2. Let us fix n ∈ N and consider ϕn : RN → [0, 1] a smooth positive mollifier radially symmetric around xn such that (1) ϕn = 1 on A3n , (2) ϕn = 0 out of A2n , (3)
ϕn  Cσn . By (2) we have ϕn = 0 and ∇ϕn = 0 on ∂A2n , and so, integrating by parts, by (1) we get 

RN

−
un un ϕn =



A2n

|∇un |2 ϕn +



A2n

∇un · ∇ϕn un

1 2 ∇ un · ∇ϕn |∇un | +
2 A2n A3n   1 = |∇un |2 −
ϕn u2n . 3 2 A2n An 

2





Therefore, being un solution to (EI), by Proposition 2.1 and (3), we have    

1 2∗
ϕn u2n  C(1 + σn )A2n . |∇un |  |un | + Aun ϕn + 2 3 2 2 An An An



2



Since σn
1 for n large, one has (2.11).

(3.21) 

3.2. Decay estimates at drift points The purpose of this second part is to establish the decay estimates and integral bounds, concerning bounded controlled sequences contained in Propositions 2.3, 2.4 and 2.6. 3.2.1. Uniform estimates. The first step is proving a lemma that allows to obtain an uniform upper bound on the values of the Laplacian on a controlled sequence. L EMMA 3.4. Let a(x) satisfy (a1 ) and (a2 ). Let (un )n∈N be a controlled sequence bounded in H 1 (RN ). Then (un )n∈N is bounded in L∞ (RN ). P ROOF. By (a1 ) and (a2 ), there exist a constant a˜ ∈ (0, a∞ ) and a positive function c(x) ∈ C0 (RN ) such that a(x)
a˜ − c(x) ∀x ∈ RN . Therefore un weakly solves p−1

−
un + au ˜ n  un

+ c(x)un

in RN ,

Multiplicity techniques for problems without compactness

565

moreover, by the maximum principle, for any weak positive solution vn ∈ H 1 (RN ) to p−1

−
v + av ˜ = un

+ c(x)un

in RN ,

(3.22)

the relation un (x)  vn (x) in RN

(3.23)

holds. Now, let us consider a sequence (vn )n∈N , vn ∈ H 1 (RN ), such that for all n ∈ N, vn solves (3.22). By (3.23), the claim follows proving that (|vn |∞ )n∈N is bounded. ∗ p−1 Since un ∈ H 1 (RN ) and c(x) ∈ C0 (RN ), we can assume un + c(x)un ∈ L2 /(p−1) , ∗ /(p−1) ∗ /(p−1) 2,2 N 2,2 N so by regularity results vn ∈ W (R ). Now, the space W (R ) embeds p−1 1 2 2∗ 2∗ q ˆ N ∗′ continuously in L (R ), where qˆ = 2∗ − N and, since p−1 > 2∗ −1 = N2N +2 = 2 , qˆ N 2∗′ ∗ q ˆ N qˆ > N −2·2∗′ = 2 . Then, by (3.23), un ∈ L (R ) with 2∗ > 1, and  p−1  |un |qˆ  |vn |qˆ  k1 un + c(x)un 2∗ /(p−1) < k2 .

By iterating the same argument, we gradually increase the regularity properties of un and vn , obtaining also uniform bounds to the norms in the respective spaces. After a finite number of steps we obtain vn ∈ W 2,q˜ (RN ) with q˜ > N2 and vn W 2,q˜ < k3 , k3 not depending on n. Then the Sobolev embedding theorem gives vn ∈ C 0,µ (RN ) for some µ ∈ (0, 1), and vn C 0,µ (RN ) < k4 . This last relation with the L2 summability allows to obtain an L∞ uniform bound on (vn )n∈N and, in turn, on (un )n∈N as desired.  C OROLLARY 3.1. Let (un )n∈N and a(x) be as in Lemma 2.3. Then there exists a constant c1 > 0 such that for all n ∈ N, the relation −
un  c1

(3.24)

weakly holds. The proof of Proposition 2.3 is carried out through some estimates, on bounded controlled sequences, proved in a slightly general setting. In order to do this we introduce the following definition. D EFINITION 3.2. Given a sequence of functions (un )n∈N , un ∈ H 1 (RN ), and a sequence (xn )n∈N , xn ∈ RN , we say that (xn )n∈N is a sequence of drift points for (un )n∈N if un (· − xn ) ⇀ 0

 weakly in H 1 RN .

(3.25)

566

S. Solimini

R EMARK 3.2. If un ⇀ 0, a sequence of points (xn )n∈N is a drift sequence if and only if σn (xn ) → +∞ and, since σn (xn )  |x|, (xn )n∈N is unbounded. In general, σn (xn ) → +∞ is equivalent to (xn )n∈N being an unbounded drift sequence. The following lemma guarantees that the values that a controlled bounded sequence takes around the drift points xn of a sequence are small. L EMMA 3.5. Let a(x) satisfy (a1 ) and (a2 ) and let (un )n∈N be a controlled sequence bounded in H 1 (RN ). Let (xn )n∈N be a drift points sequence for (un )n∈N and let δn = σn (xn ). Then for all h ∈ (0, 1), lim

n→+∞ B

sup un (x) = 0.

(3.26)

hδn (xn )

P ROOF. We argue by contradiction and we assume that there exist real numbers h ∈ (0, 1), η > 0 and a sequence (yn )n∈N , yn ∈ Bhδn (xn ), such that for large n, un (yn ) >

3

4 1 sup un (x) − > η. n Bhδn (xn )

The above relation, combined with (3.24), allows to conclude that, for large n and ρ small enough,  −

Bρ (yn )

un dx =

1 |Bρ (yn )|



η un dx > , 2 Bρ (yn )

where |Bρ (yn )| denotes the Lebesque N -dimensional measure of Bρ (yn ). Hence un (· − yn ) ⇀ v = 0 as n → +∞. This is impossible because, by the choice of δn , h and yn and by Definition 3.2 un (· − yn ) ⇀ 0 in H 1 (RN ).  Next lemma contains the key estimate for proving Proposition 2.3. L EMMA 3.6. Let a(x) satisfy (a1 ) and (a2 ) and let (un )n∈N be a controlled sequence bounded in H 1 (RN ). Let (xn )n∈N be an unbounded drift points sequence for (un )n∈N . √ Then, for all α ∈ (0, a∞ ), there exists a constant cˆα > 0 such that for all n, un (xn )  cˆα e−ασn (xn ) .

(3.27)

√ P ROOF. Let α ∈ R, 0 < α < a∞ , be fixed, and let us choose h ∈ ( √αa , 1) and α¯ ∈ ∞ √ (α, a∞ h). Then, by using Lemma 3.5, we obtain that, for any n large enough, un weakly satisfies p−1


un
a(x)un − un

> α¯ 2 h−2 un
0 in Bhδn (xn ),

(3.28)

Multiplicity techniques for problems without compactness

567

where δn = σn (xn ). Thus, since hδn > 1 for large n, we have un (xn ) 



un dσ

∂Br (xn )

∀r: 0 < r  1,

(3.29)

and we deduce un (xn ) 

 1 ' 0

∂Br (xn )

(  un dσ dr =

un dx.

B1 (xn )

So, in order to obtain (3.27) for large n, it is enough to show that a constant c¯α > 0 exists such that 

B1 (xn )

un dx  c¯α e−αδn .

(3.30)

To do this, let us consider the functions vn (ρ) =



un dx,

Bρ (xn )

wn (ρ) =

(hδn )N ωN αρ/ e ¯ h, ¯ n eαδ

where ωN is the Lebesgue measure of the unitary ball in RN , and let us remark that vn (1) is just the left-hand side of (3.30), while wn (1) =

N √ (hδn )N ωN α/ ¯ h a∞ δn e  hω e  c¯α e−αδn N ¯ n ¯ n eαδ eαδ

for n large enough. So (3.27) follows, by proving vn (1)  wn (1) and taking into account ¯ (3.27) is obviously true for a suitable choice of the that for any finite set un (xn ), n < n, constant cˆα . Let us then show that for n large vn (ρ)  wn (ρ)

∀ρ ∈ [0, hδn ].

First, let us observe that vn (0)  wn (0)

∀n ∈ N,

and that for n large, by Lemma 3.5,   vn (hδn )  Bhδn (xn ) sup un (x)  ωN (hδn )N = wn (hδn ). Bhδn (xn )

(3.31)

568

S. Solimini

Now, if for some point in [0, hδn ] (3.31) were false, then the function (vn − wn )(ρ) should have a maximum point, ρ¯n ∈ (0, hδn ), for which (vn − wn )(ρ¯n ) > 0 and, of course, v ′′ (ρ¯n ) − w ′′ (ρ¯n )  0. Let us show that this is impossible. Indeed, since (   ρ ' vn (ρ) = un dσ dr, un (x) dx = 0

Bρ (xn )

∂Br (xn )

we have vn′ (ρ) = moreover,  −

∂Bρ (xn )

d vn (ρ) = dρ

un dσ =



un dσ ,

∂Bρ (xn )

1 N ωN ρ N −1



∂Bρ (xn )

un dσ =

vn′ (ρ) N ωN ρ N −1

and, by using divergence theorem,  −

∂Bρ (xn )

un dσ = =



ρ

0



ρ

0

'

( ∂ un dσ dr ∂Br (xn ) ∂ν ' ( 1
u dx dr. n N ωN r N −1 Br (xn ) 1 N ωN r N −1

So d 1 vn′ (ρ) = N −1 dρ N ωN ρ N ωN ρ N −1




un dx,

Bρ (xn )

from which, using (3.28), we obtain vn′′ (ρ) vn′ (ρ) + (1 − N ) ρN ρ N −1  ′  1 d vn (ρ) α 2 h−2 = =
u dx
vn (ρ). n dρ ρ N −1 ρ N −1 Bρ (xn ) ρ N −1 Hence, taking into account that vn′ (ρ) > 0 and N > 1, vn′′ (ρ)
α 2 h−2 vn (ρ)

∀ρ ∈ (0, hδn )

(3.32)

follows. Let now ρ¯n ∈ (0, hδn ) be a maximum point for (vn − wn )(ρ) for which (vn − wn )(ρ¯n ) > 0 then by (3.32), we get

 vn′′ (ρ¯n ) − wn′′ (ρ¯n )
α 2 h−2 vn (ρ¯n ) − wn (ρ¯n ) > 0,

Multiplicity techniques for problems without compactness

569



and we are in contradiction.

P ROOF OF P ROPOSITION 2.3. Arguing by contradiction, we assume that there are √ α ∈ (0, a∞ ), (xn )n∈N and a sequence of integers kn ∈ N such that ukn (xn ) > ne−ασkn (xn ) or, by replacing (un )n∈N by the subsequence (xkn )n∈N , un (xn ) > ne−ασn (xn ) . By Lemma 3.4 we get σn (xn ) → +∞ so we get in contradiction to Lemma 3.6 whose  assumptions are fulfilled by (xn )n∈N . P ROOF OF P ROPOSITION 2.6. By using Proposition 2.3, we deduce that, for n large √ enough and α ∈ (0, a∞ ), 

Sn

p

(un ) dx  cα  cα



Sn

e−αpσn (x) dx

k  

i

i=0 Sn

 cα k



+∞

γˆ 4k |tn |

e−αp|x−tn | dx ˜ n| e−αpt t N −1 dt  ce ˜ −α|t .



3.2.2. Gradient estimates. Also in this case, a Caccioppoli-type inequality allows to pass to an integral estimate of the gradient. P ROOF OF P ROPOSITION 2.4. For any fixed n ∈ N, let ϕn ∈ C ∞ (RN , [0, 1]) be a function fulfilling the following conditions:  on S1,n ,  (i) ϕn = 1 (ii) supp(ϕn ) ⊂ S2,n ,  (iii)
ϕn  C, C ∈ R.

(3.33)

Since un weakly solves (EI) and ϕn = 0 in RN \ S2,n , we have   (−
un )(un ϕn ) = (−
un )(un ϕn ) RN

S2,n





RN

=





p un − a(x)u2n ϕn

S2,n

p  un − a(x)u2n ϕn .

(3.34) (3.35)

570

S. Solimini

On the other hand, taking into account that by (3.33)(ii), ϕn = 0 and ∇ϕn = 0 on ∂S2,n and using (3.33)(i), we get 

S2,n

(−
un )(un ϕn ) =



S2,n

|∇un |2 ϕn +



S2,n

(∇un · ∇ϕn )un

  1 ∇ u2n · ∇ϕn 2 S1,n S2,n   1 = |∇un |2 − (
ϕn )u2n . 2 S2,n S1,n




|∇un |2 +



(3.36)

So, inserting (3.34) in (3.36), using (3.33)(iii) and taking into account that, if n is large enough, S2,n ⊂ Sn and a(x) > 0 on Sn , we deduce for large n, 

S1,n



p  1 (
ϕn )u2n un − a(x)u2n ϕn + 2 S2,n S2,n   1 p un ϕn +  (
ϕn )u2n 2 Sn Sn   1 p un + C  u2n . 2 Sn Sn

|∇un |2 



Then, applying Proposition 2.6, taking also into account that for any finite set of indexes (2.22) is true for a suitable choice of the constant c∗ , we obtain the thesis.  4. Multiplicity results This section is devoted to the proof of the two multiplicity theorems (Theorems 1.2 and 1.4). The proofs will be achieved by using the compactness results proved in Section 2 and wellknown variational tools, employed for the approximating problems, based on the use of Krasnoselskii genus. Traditionally this approach is used for searching constrained critical points, however we shall use a recent variant of this method which works with unconstrained min–max classes [43]. This will make working with several functionals at the same time easier, will simplify the use of the Morse index and will let us make a final application in the next section in a case in which the constrained approach is made difficult by a lack of regularity of the constraint manifold. So we shall begin by introducing the classical Krasnoselskii genus, then we shall pass to the variant introduced in [43], we shall show how the min–max approaches on the natural constraint V are equivalent to an unconstrained min–max, we shall introduce the double natural constraint W showing that it enjoys the same property and finally we shall give an estimate on the Morse index. After this introductory parts, in the last two parts of this section we shall give the proof of the two multiplicity theorems.

Multiplicity techniques for problems without compactness

571

4.1. Krasnoselskii genus We recall some well-known facts about the Krasnoselskii genus in view of its application to semilinear elliptic equations of the type addressed here, moreover, we introduce some generalizations of those concepts introduced in [43] which will be useful in the following. We refer to [26] and [46] for a more comprehensive treatment of the subject. Throughout this subsection we shall denote by E any given Banach space. Let A ⊂ E be a symmetric subset of E, i.e., A = −A, we define Λk (A) as the space of Krasnoselskii test maps on A of dimension n as follows 

 Λk (A) = ϕ ∈ C A, Rn | ϕ(x) = −ϕ(−x) .

We shall call Krasnoselskii genus of A the number γ (A) so defined  γ (A) = inf n ∈ N | ∃ϕ ∈ Λk (A): 0 ∈ / ϕ(A) .

We notice that, from the above definitions, for any given ϕ ∈ Λk (A), if k < γ (A) then 0 ∈ ϕ(A), whereas if k
γ (A) then there exists ϕ ∈ Λk (A) such that 0 ∈ / ϕ(A). The main properties of the Krasnoselskii genus, which will be useful for the subsequent arguments, rely on the following statements. P ROPOSITION 4.1. Let A be any given symmetric subset of a Banach space and let η : A → E be a given odd map. Then γ (η(A))
γ (A). P ROOF. Let k < γ (A) and let ϕ ∈ Λk (η(A)). Since ϕ ◦ η ∈ Λk (A), then we have 0 ∈ ϕ(η(A)) and therefore γ (η(A)) > k. The thesis follows from the arbitrariness of k.  P ROPOSITION 4.2. Let S ⊂ E be any closed subspace of co-dimension k ∈ N. If A ⊂ E is any symmetric subset such that γ (A) > k, then A ∩ S = ∅. P ROOF. Let S ⊥ be the orthogonal complement of S in E, which is isomorphic to Rk , and let P : E → S ⊥ be the orthogonal projection map. Then P ∈ Λk (A) and 0 ∈ P (A) so, since P −1 (0) = S, we have that S ∩ A = ∅.  P ROPOSITION 4.3. γ (S k−1 ) = k. P ROOF. Let ϕ ∈ Λk−1 (S k−1 ), by the Borsuk theorem we have 0 ∈ ϕ(S k−1 ) and so γ (S k−1 ) > k − 1. On the contrary, the canonical injection i : S k−1 → Rk belongs to Λk (S k−1 ) and 0 ∈ / i(S k−1 ), therefore γ (S k−1 )  k.  The next proposition states a trace property for the Krasnoselskii genus. P ROPOSITION 4.4. Let A ⊂ E be any given symmetric subset and let ϕ ∈ Λk (A), with k < γ (A). Then γ (ϕ −1 (0))
γ (A) − k.

572

S. Solimini

P ROOF. Firstly, let us notice that ϕ −1 (0) is a symmetric set. We take h < γ (A) − k and ψ ∈ Λh (ϕ −1 (0)). By virtue of the Tietze extension theorem, ψ may be extended to an ¯ ∈ Rk+h , with odd map ψ¯ ∈ C(A, Rk ). We introduce the map ϕ × ψ¯ : x → (ϕ(x), ψ(x)) ¯ k + h < γ (A). Since ϕ × ψ¯ is a Krasnoselskii test map we have that 0 ∈ (ϕ × ψ)(A) and this means that 0 ∈ ψ(ϕ −1 (0)). So γ (ϕ −1 (0)) > h and by arbitrariness of h we get the thesis. 

4.2. Genus of a symmetric set For any k ∈ N, we adopt the following notation: Qk = [−1, 1]k , F±i = {x ∈ Qk | xi = ±1}. Given n, k ∈ N, for every x ∈ Rn+k , we shall split x as x = (x0 , x1 , x2 , . . . , xk ) with x0 ∈ Rn and xi ∈ R for i = 1, . . . , k. Analogously, if ϕ(x) ∈ Rn+k , we shall write ϕ(x) = (ϕ0 (x), ϕ1 (x), . . . , ϕk (x)) with ϕ0 (x) ∈ Rn and ϕi (x) ∈ R for i = 1, . . . , k. Sometimes, we shall also use the notation x = (x0 , x ′ ), ϕ(x) = (ϕ0 (x), ϕ ′ (x)) instead of the previous one, by assuming x ′ , ϕ ′ (x) ∈ Rk . We shall set Bn = {x ∈ Rn | |x|  1}, S n = ∂Bn+1 = {x ∈ Rn+1 | |x| = 1} and we shall denote by E any given Banach space. For any given k ∈ N, we set Ik = {1, 2, . . . , k} and we denote by Pk the set of all the involutive permutations on Ik , i.e., π ∈ Pk if and only if π : Ik → Ik and π ◦ π = id. Given any π ∈ Pk , we introduce the map πˆ : Rk → Rk so defined πˆ (x1 , x2 , . . . , xk ) = (xπ(1) , xπ(2) , . . . , xπ(k) ). We have

 πˆ π(x) ˆ = πˆ (xπ(1) , xπ(2) , . . . , xπ(k) ) = (xπ(π(1)) , xπ(π(2)) , . . . , xπ(π(k)) ) = x,

which means πˆ ◦ πˆ = id. Furthermore, given any Banach space E, let us define the map πE : E × Rk → E × Rk such that πE : (x0 , x ′ ) → (−x0 , π(x ˆ ′ )). It is easily seen that πE is involutive too, indeed

 



 πE πE (x) = πE −x0 , πˆ x ′ = x0 , πˆ πˆ x ′ = x0 , x ′ = x

and so we also have πE ◦ πE = id. We shall set π˜ = πE if E = Rn . Now, let A ⊂ E be any symmetric subset and let ϕ : A × Qk → Rn+k and π ∈ Pk be given, we set

 ϕπ = π˜ ◦ ϕ ◦ πA : x → π˜ ϕ πE (x) ,

where πA : A × Qk → A × Qk is the restriction of πE . We shall refer to ϕπ as to the π -symmetric of ϕ. We shall say that ϕ is π -symmetric if ϕ = ϕπ and that ϕ is symmetric if there exists π such that ϕ = ϕπ . We note that (ϕπ )π = π˜ ◦ (π˜ ◦ ϕ ◦ πA ) ◦ πA = ϕ. R EMARK 4.1. We observe that if π = id ϕ is π -symmetric if (S1) ϕ0 (−x0 , x ′ ) = −ϕ0 (x0 , x ′ ) ∀x ∈ A × Qk , (S2) ϕ ′ (−x0 , x ′ ) = ϕ ′ (x0 , x ′ ) ∀x ∈ A × Qk .

Multiplicity techniques for problems without compactness

573

D EFINITION 4.1. Let A ⊂ E be any given symmetric subset, for any k ∈ N we shall say that a symmetric function ϕ ∈ C(A × Qk , Rn+k ) is a test function of dimension n for A if the following condition holds (T) ±ϕi (x)
0 ∀x ∈ A × F±i , ∀i = 1, . . . , k. We shall denote by Λ∗n (A) the set of the n-dimensional test functions for A, obtained for any value of k. Then we are ready to define the genus of a set A as the number    / ϕ(A × Qk ) . γ ∗ (A) = inf n ∈ N  ∃ϕ ∈ Λ∗n (A)  s.t. 0 ∈

Firstly, we remark that, since Λ∗n (A) is a larger set than the set Λn (A) of the test maps related to the Krasnoselskii genus (indeed the last one coincides with the subset of the former one which contains the test functions constructed by taking k = 0), we have that, in general, γ ∗ (A)  γ (A). By definition, given any ϕ ∈ Λ∗n (A) with n < γ ∗ (A), 0 ∈ ϕ(A). In view of proving the analogous statement of Propositions 4.1–4.4 for the genus γ ∗ , we prove some useful lemmas stating some properties of the test functions. L EMMA 4.1. Let ϕ : A × Qk → Rn+k satisfy (T) and let π ∈ Pk be given, then ϕπ also satisfies (T). ± P ROOF. Let x = (x0 , x ′ ) ∈ A × Fi± for i = 1, . . . , k. Then πˆ (x ′ ) ∈ Fπ(i) and, since ϕ satisfies (T), we have







±ϕπ(i) πE x0 , x ′ = ±ϕπ(i) −x0 , πˆ x ′
0.

By the previous definitions we know that ϕπ(i) = (π˜ ◦ ϕ)i and so ±(π˜ ◦ ϕ)i (πE (x))
0, that is, ±(ϕπ )i (x)
0, as stated in the thesis.  L EMMA 4.2. Let A ⊂ E be any given symmetric subset. Then every test function ϕ ∈ Λ∗n (A) can be extended to a map in Λ∗n (E). P ROOF. Firstly, by virtue of the Tietze–Dugundji theorem (see [22]) we take an extension of the components ϕi on E × Fi± valued in R± keeping the sign property (T), then we extend them and ϕ0 continuously on all of E × Qk so getting a map ϕ¯ : E × Qk → Rn+k which, obviously, satisfies (T). By Lemma 4.1, the map ϕ¯π , which is an extension of ϕπ , satisfies (T) and the same happens for ϕs = 21 ϕ¯ + 12 ϕ¯ π : E × Qk → Rn+k , where π ∈ Pk is such that ϕ is π -symmetric. Obviously, ϕs is symmetric since 1 1 1 1 (ϕs )π = ϕ¯π + (ϕ¯π )π = ϕ¯ π + ϕ¯ = ϕs , 2 2 2 2 so ϕs ∈ Λ∗n (E). If x ∈ A × Qk then ϕ(x) = ϕπ (x), hence ϕ(x) = ϕ(x) ¯ = ϕ¯ π (x) and so ϕ(x) = ϕs (x). Therefore ϕs extends ϕ to all of E.  The property in Proposition 4.1 also holds for γ ∗ .

574

S. Solimini

P ROPOSITION 4.5. Let A ⊂ E be any given symmetric subset and let η : A → E be a given odd continuous map. Then γ ∗ (η(A))
γ ∗ (A). P ROOF. Let n < γ ∗ (A), ϕ ∈ Λn (η(A)) and let η(x ¯ 0 , x ′ ) = (η(x0 ), x ′ ), η¯ : A × Qk → η(A) × Qk . Given π ∈ Pk , we have





 πE η¯ x0 , x ′ = −η(x0 ), πˆ x ′ = η(−x0 ), πˆ x ′



 = η¯ −x0 , πˆ x ′ = η¯ πE x ′

and so πη(A) ◦ η¯ = η¯ ◦ πA . We take ϕ¯ = ϕ ◦ η¯ : A × Qk → Rn+k , then ¯ ϕ¯π = π˜ ◦ ϕ¯ ◦ πA = π˜ ◦ ϕ ◦ η¯ ◦ πA = π˜ ◦ ϕ ◦ πη(A) ◦ η¯ = ϕπ ◦ η. Therefore, if we fix π such that ϕ is π -symmetric, that is, ϕ = ϕπ , then ϕ¯ π = ϕ ◦ η¯ = ϕ, ¯ hence also ϕ¯ is symmetric. Since η(A ¯ × Fi± ) = η(A) × Fi± , we deduce that ϕ¯ satisfies (T). Thus ϕ¯ ∈ Λ∗n (A), hence



 0 ∈ ϕ(A ¯ × Qk ) = ϕ η(A ¯ × Qk ) = ϕ η(A) × Qk .

It follows that n < γ ∗ (η(A)) by the arbitrariness of ϕ and, consequently, γ ∗ (A)  γ ∗ (η(A)) by the arbitrariness of n.  Proposition 4.2 directly applies to γ ∗ since γ ∗ (A)  γ (A). In order to prove the analogous of Proposition 4.3 we need a topological lemma which states a more general property than the Borsuk theorem. T HEOREM 4.1. Let f : Bn × Qk → Rn+k be a continuous map symmetric on ∂Bn × Qk (i.e., such that for some π ∈ Pk , f (x) = fπ (x) ∀x ∈ S n−1 × Qk ) and assume that (T) holds for ϕ = f and A = Bn . Then there exists x ∈ Bn × Qk such that f (x) = 0. P ROOF. The first step consists in introducing a suitable change of variables which will allow to deal with the symmetry properties involved in the most convenient way. For given k ∈ N and π ∈ Pk , let us introduce the sets  I0 = i ∈ Ik | i = π(i) ,

 I1 = i ∈ Ik | i < π(i) .

We note that, if ki = ♯Ii , k0 + 2k1 = k. Let us introduce the functions p1 and p2 , both defined on Rk as p1 (x) = (xi + xπ(i) )i∈I1 + (xi )i∈I0 ,

p2 (x) = (xi − xπ(i) )i∈I1 .

We have Rk ∼ = p1 (Rk ) ⊕ p2 (Rk ). Then for every x ∈ Bn × Qk , we set

 x 1 = x 0 , p2 x ′ ,

 x 2 = p1 x ′ ,

Multiplicity techniques for problems without compactness

575

and so we have x = (x0 , x ′ ) ∼ = (x1 , x2 ). Analogously, for any f = (f0 , f ′ ) ∈ Rn+k , we define the functions

 f1 = f0 , p2 ◦ f ′ ,

f2 = p1 ◦ f ′ ,

and so f ∼ = (f1 , f2 ). Now we observe that x1 = 0 if and only if x0 = 0 and, for every i ∈ Ik , ˆ ′ )) = π(x) ˜ and that, consequently, f π -symmetric xi = xπ(i) , that (−x1 , x2 ) ∼ = (−x0 , π(x means

  f (−x1 , x2 ) = π˜ ϕ(x1 , x2 ) = −f1 (x1 , x2 ), f2 (x1 , x2 ) .

(4.1)

We assume by contradiction that

  k

+ 

n−1  * −  0∈ / f ∂(Bn × Qk ) = f S Bn × Fi ∪ Fi . × Qk ∪ i=1

In such a case, we shall show that the topological degree of f in zero is different from zero, that is, deg(Bn × Qk , f, 0) = 0. The first step in this direction consists in forcing the assumption (T) to be satisfied with strict inequalities, by adding to f ′ the function εx ′ with ε > 0 suitably small in order to keep the value of the topological degree. We just remark that the function g : (x0 , x ′ ) → (0, x ′ ) is π -symmetric, as it can be easily seen. Then, through a linear homotopy, we can pass from f to 21 (f +fπ ), which we will continue to denote by f . By Lemma 4.1 we know that the symmetrization does not change the degree since f is not modified on ∂Bn ×Qk and fi and (fπ )i have a fixed sign on Bn ×Fi± . By a standard perturbation argument we can also assume f ∈ C 1 (Bn × Qk , Rn+k ). We set X = {x ∈ Rn+k | x1 = 0}. We shall introduce further modifications of f which make f −1 (0) ∩ X contain only regular points. Firstly we know that, by the oddness of f1 with respect to the variable x1 stated in (4.1), f1 = 0 on X and so the zeros of f on X are the zeros of f2 . Then we observe that, since after the previous symmetrization f2 is even with respect to the variable x1 , so the partial Jacobian matrix Jx1 f2 (x) is identically zero for every x ∈ X. Then, for any x ∈ X, by Laplace rule we get |Jf (x)| = |Jx1 f1 (x)||Jx2 f2 (x)|. We can force |Jx2 f2 | = 0 on every x ∈ X such that f2 (x) = 0 by subtracting from f2 a small regular value h ∈ p1 (Rk ), given by the Sard theorem. Note that this perturbation does not affect the symmetry properties of f and, if h is taken sufficiently small, does not change the value of the topological degree. In particular, we get that f −1 (0) ∩ X is a finite set and therefore the set L which contains all the eigenvalues of Jx1 f1 (x) in the points of f −1 (0) ∩ X is also finite. Then we can force the determinant |Jx1 f1 (x)| to be different from zero in such points by adding to f1 the function −λx1 with λ ∈ R \ L. Again, this new perturbation preserves the symmetry properties of f and, if λ is sufficiently small, does not change the value of deg(Bn × Qk , f, 0) and the zeros of f2 remain of course the same. So we can be sure that f has only regular zeros on X. Now, since f2 satisfies the hypotheses of the Miranda theorem (see [33]) on X ∼ = Qk0 +k1 , we can state that f −1 (0) ∩ X is composed by an odd number of regular zeros. By continuity, we have that there exists a small ε > 0 such that the closed tubular neighborhood

576

S. Solimini

Xε = {x ∈ Bn × Qk | d(x, X)  ε} contains only regular zeros. In order to force 0 to be a regular value for f we have to deal with the set (Bn × Qk ) \ Xε and, to this aim, we argue as follows. ε i i Let A± i = {x ∈ Bn × Qk | ±x1
n }, where, for i = 1, . . . , n + k1 , x1 is the component of x1 of index i. One can easily see that (Bn × Qk ) \ Xε ⊂

n+k *1 i=1



+ Ai ∪ A− i .

We are going to perform a new perturbation of f , which keeps the symmetry properties and is too small to change the topological degree or to introduce singular zeros in Xε , in order to exclude the presence of singular zeros of f also on A± 1 . Let S be the set of singular values of f . By the Sard theorem S ∪ π(S) ˜ is a negligible set, so we can take h1 ∈ Rn+k \ (S ∪ π˜ (S)) arbitrarily small. Let g : R → [0, 1] be a smooth function such that g(x) = 0 for x  0 and g(x) = 1 for x
1. Let ψ : Rn → Rn+k be defined as ψ(x) = g( nε x11 )h1 +g(− nε x11 )π˜ (h1 ). One easily sees that ψ(x) = h1 if x ∈ A+ ˜ 1 ) if 1 and ψ(x) = π(h ¯ : x → f (x)−ψ(x) has no singular zeros on A+ ∪A− . Moreover, x ∈ A− . So the function f 1 1 1 f¯ keeps the symmetry properties of f . The value of the degree and the regularity of the zeros in Xε are preserved by stability, provided h1 is taken suitably small. We proceed in this construction through n + k1 steps, from i = 1 to i = n + k1 . Thanks to the stability property of the regular points, we are sure that at each step the regularity gained  + − at the previous step on Xε ∪ i−1 j =1 (Aj ∪ Aj ) is kept, provided the perturbation term hi given by the Sard theorem is chosen sufficiently small. Therefore, we can conclude that we have regularity everywhere on Bn × Qk and so 0 is a regular value for f . Finally, we know that f −1 (0) ∩ X is made by an odd number of points and, since f −1 (0) = π(f ˜ −1 (0)), −1 f (0) \ X is made by an even number of points, indeed π(x) ˜ = x for x ∈ / X. Then we can conclude that deg(Bn × Qk , f, 0) is odd and so we get the thesis.  R EMARK 4.2. It is worth to notice that the previous theorem reduces to the Borsuk theorem when k = 0 and to the Miranda theorem when n = 0 and π = id. If it is easy to see that Miranda theorem can be deduced from the Borsuk theorem, nevertheless reconducing the above statement to the Borsuk theorem does not seem to be an obvious task. P ROPOSITION 4.6. γ ∗ (S n ) = n + 1. P ROOF. We know that γ ∗ (S n )  γ (S n )  n + 1. On the other hand, let ϕ ∈ Λ∗n (S n ) and n = {(x , . . . , x n n let S+ 1 n+1 ) ∈ S | xn+1
0}. S+ is homeomorphic to Bn and so, by Theon rem 4.1, there exists x ∈ S+ × Qk such that ϕ(x) = 0. So n < γ ∗ (S n ).  Also the trace property in Proposition 4.4 is enjoyed by γ ∗ . P ROPOSITION 4.7. Let A ⊂ E be any given symmetric subset and let ϕ ∈ Λk (A), with k < γ ∗ (A). Then γ ∗ (ϕ −1 (0))
γ ∗ (A) − k.

Multiplicity techniques for problems without compactness

577

P ROOF. Let k  n < γ ∗ (A) and ψ ∈ Λ∗n−k (ϕ −1 (0)), ψ : ϕ −1 (0) × Qh → Rn−k+h . By Lemma 4.2, ψ can be extended to ψ¯ defined on all of E (and so on A) and such that ψ¯ ∈ Λ∗n−k (A). Let us consider the function ϕ × ψ¯ : A × Qh → Rn+h defined as





  ¯ ϕ × ψ¯ (x) = ϕ(x0 ), ψ(x) = ϕ(x0 ), ψ¯ 0 (x) , ψ¯ ′ (x) .

It is easily seen that ϕ × ψ¯ satisfies (T), since the components involved in such condition only belong to ψ¯ . Moreover, if π ∈ Ph then

 



 π˜ ϕ × ψ¯ πE (x) = π˜ ϕ(−x0 ), ψ¯ 0 πE (x) , ψ¯ ′ πE (x)

  = −ϕ(−x0 ), −ψ¯ 0 πE (x) , πˆ ψ¯ ′ πE (x)

  ′  = ϕ(x0 ), π˜ ◦ ψ¯ 0 πE (x) , π˜ ◦ ψ¯ πE (x)

 = ϕ × ψ¯ π (x).

¯ π = ϕ × ψ¯ π , hence since ψ¯ is symmetric, then also ϕ × ψ¯ satisfies Therefore (ϕ × ψ) ¯ × Qh ), that is, 0 ∈ the same condition. Thus ϕ × ψ¯ ∈ Λn (A) and then 0 ∈ (ϕ × ψ)(A ¯ −1 (0) × Qh ). ψ(ϕ By the arbitrariness of ψ¯ , we get n − k < γ ∗ (ϕ −1 (0)) and by the arbitrariness of n, the thesis follows.  A different trace property, which actually is the main motivation for passing to the present variant of the notion of genus, can be now also proved. P ROPOSITION 4.8. Let A ⊂ E be any symmetric subset and let σ : A × Qk → E and ϕ : E → Rk be given. If there exists π ∈ Pk such that (a) ∀x ∈ A × Qk : σ (πE (x)) = −σ (x), (b) ∀u ∈ E: ϕ(−u) = π(ϕ(u)), ˆ (c) ∀x ∈ A × Fi± : ±ϕi (σ (x))
0 ∀i = 1, . . . , k, then γ ∗ (σ (A × Qk ) ∩ ϕ −1 (0))
γ ∗ (A). P ROOF. Firstly we observe that conditions (a) and (b) allow to state respectively that σ (A × Qk ) and ϕ −1 (0) are symmetric subsets of E and so the assertion makes sense. We fix n < γ ∗ (A) and a test function ψ of dimension n defined on (σ (A × Qk ) ∩ ϕ −1 (0)) and extended by Lemma 4.2 on E. Then ψ : E × Qh → Rn+h satisfies (T) and is symmetric for some π1 ∈ Ph . Let us define ρ ∈ Pk+h such that ρ(i) = π1 (i) if i  h and ρ(i) = h + π(i − h) if i > h. In this way, if x ′ ∈ Rh+k is decomposed as x ′ = (z′ , y ′ ) with ˆ ′ ) = (πˆ 1 (z′ ), πˆ (y ′ )). We define ϑ : A × Qh+k → Rn+h+k z′ ∈ Rh and y ′ ∈ Rk , we have ρ(x by setting



    ϑ x0 , x ′ = ϑ x0 , z ′ , y ′ = ψ σ x0 , y ′ , z ′ , ϕ σ x0 , y ′ ,

that is, ϑ0 = ψ0 and ϑ ′ = (ψ ′ , ϕ). We are going to prove that ϑ ∈ Λ∗n (A). Since ψ satisfies (T), we have that, for i  h, ±ϑi = ±ψi
0 if z′ ∈ Fi± , whereas condition (c) states

578

S. Solimini

that ±ϑh+i = ±ϕi
0 if y ′ ∈ Fi± and so ϑ satisfies (T). Moreover, by (a), (b) and the symmetry of ψ ,



  ϑ ρE (x) = ϑ −x0 , πˆ 1 z′ , πˆ y ′

 



= ψ σ −x0 , πˆ y ′ , πˆ 1 z′ , ϕ σ −x0 , πˆ y ′   

= ψ σ πE x0 , y ′ , πˆ 1 z′ , ϕ σ πE x0 , y ′   

= ψ −σ x0 , y ′ , πˆ 1 z′ , ϕ −σ x0 , y ′   



= ψ (π1 )E σ x0 , y ′ , z′ , πˆ ϕ σ x0 , y ′

   = π˜ 1 ψ σ x0 , y ′ , z′ , πˆ ϕ σ x0 , y ′

 

  

= −ψ0 σ x0 , y ′ , z′ , πˆ 1 ψ ′ σ x0 , y ′ , z′ , πˆ ϕ σ x0 , y ′

    

= −ψ0 σ x0 , y ′ , z′ , ρˆ ψ ′ σ x0 , y ′ , z′ , ϕ σ x0 , y ′



= −ϑ0 (x), ρˆ ϑ ′ (x)

 = ρ˜ ϑ(x) . Then ϑ is symmetric and so ϑ ∈ Λ∗n (A) as claimed above and, since n < γ ∗ (A), 0 ∈ ϑ(A × Qh+k ). This means that there exist x0 ∈ A, y ′ ∈ Qk and z′ ∈ Qh such that ψ(σ (x0 , y ′ ), z′ ) = 0 and ϕ(σ (x0 , y ′ )) = 0. So 0 ∈ ψ(((σ (A × Qk ) ∩ ϕ −1 (0))) × Qh ). By the arbitrariness of ψ , one gets n  γ ∗ (σ (A × Qk ) ∩ ϕ −1 (0)) and by the arbitrariness of n, one gets the thesis.  4.3. Min–max classes on the natural constraint Let I : H → R be one of the functionals related to the problem (CP) or to the problem (P), with H = H01 (Ω) or H = H 1 (RN ), respectively. Let V be the natural constraint defined in (1.16). For any fixed n we introduce the class of sets  ΓnV = A ⊂ V | A compact, γ ∗ (A)
n and denote by cn the corresponding min–max level, that is, cn = inf sup I (u). A∈ΓnV u∈A

We assume cn > 0. Such an assumption will let us find a neighborhood C − of 0 such that for every u ∈ C − , I (u) 

cn , 2

∇I (u) · u
0,

in conjunction to which we take a neighborhood of infinity C + such that for every u ∈ C + , I (u)  0 and therefore ∇I (u) · u  0.

Multiplicity techniques for problems without compactness

579

Let us observe that, given any u ∈ V, the function α → I (αu) is strictly increasing between 0 and 1, strictly decreasing after 1 and tends to −∞ as α → +∞. So two constants ε, c > 0 such that εu ∈ C − and cu ∈ C + always exist. Moreover, if A is any compact subset of V, ε and c can be uniformly fixed for u ∈ A, also ε can be fixed in maximal way and c can be fixed in minimal way such that both the values turn out to be considered as functions of A. We introduce the function σA : A × [−1, 1] → H , defined as  c−ε u, σA (u, α) = ε + (α + 1) 2 satisfying the conditions σA (u, −1) = εu,

σA (u, 1) = cu ∀u ∈ A

and the set

 ΣA = σA A × [−1, 1]

which will be named simple homothetic expansion of A. Let us notice that if A ∈ ΓnV then σA belongs to the class of continuous functions defined by   Fn = σ : A × [−1, 1] → H  A ∈ ΓnV , ∀u ∈ A: σ (−u, α) = −σ (u, α), σ (u, −1) ∈ C − , σ (u, 1) ∈ C +

and ΣA belongs to the class of sets defined by   Γn = X ⊂ H \ {0}  ∃A ∈ ΓnV ,

 ∃σ : A × [−1, 1] → H, σ ∈ Fn , s.t. X = σ A × [−1, 1] .

Let us introduce two further classes of continuous functions and sets, namely   Fn∗ = ϕ : H → R  ϕ(−u) = −ϕ(u),

ϕ(u)  0 ∀u ∈ H ∩ C − , ϕ(u)
0 ∀u ∈ H ∩ C + ,  

 Γn∗ = X ⊂ H \ {0}  X compact, X = −X, ∀ϕ ∈ Fn∗ : γ ∗ X ∩ ϕ −1 (0)
n .

L EMMA 4.3. Γn ⊂ Γn∗ .

P ROOF. Let ϕ ∈ F ∗ , A ∈ ΓnW , σ : A × Q1 → H , σ ∈ Fn be fixed. By virtue of Theorem 4.8 we have γ ∗ (σ (A × Q1 ) ∩ ϕ −1 (0))
n. By the arbitrariness of ϕ ∈ F ∗ , we get  σ (A × Q1 ) ∈ Γn∗ . L EMMA 4.4. For every A ∈ Γn∗ , we have A ∩ V ∈ ΓnV .

580

S. Solimini

P ROOF. To prove the statement it suffices to note that the function ϕ : u → ∇I (u) · u belongs to Fn∗ and hence, by the definition of Γn∗ , the thesis follows.  L EMMA 4.5. If A ∈ ΓnV then supΣA I = supA I . P ROOF. The assertion trivially follows from the definition of ΣA .



L EMMA 4.6. infA∈Γn supu∈A I (u) = infA∈Γn∗ supu∈A I (u) = cn . P ROOF. By virtue of the inclusion Γn ⊂ Γn∗ , we have inf sup I (u)  inf sup I (u).

A∈Γn∗ u∈A

A∈Γn u∈A

Moreover, for every X ∈ Γn∗ , from Lemma 4.4 we have cn  sup  sup I (u) u∈X∩V

u∈X

and so, by the arbitrariness of X, cn  infA∈Γn∗ supu∈A I (u). Finally, by Lemma 4.5 we have that, for every X ∈ ΓnV , since ΣX ∈ Γn , inf sup I (u)  sup I (u) = sup I (u).

A∈Γn u∈A

u∈ΣX

u∈X

By the arbitrariness of X in ΓnV , we have inf sup I (u)  inf sup I (u) = cn .

A∈Γn u∈A

A∈ΓnV u∈A



P ROPOSITION 4.9. Γn and Γn∗ are two admissible min–max classes. P ROOF. Since cn > 0, we can find deformations η reducing the level cn and leaving unchanged the sublevel of c2n . Obviously such deformations do not change the sets C − and C + . It follows that if σ ∈ Fn and ϕ ∈ Fn∗ then η ◦ σ ∈ Fn and so if A ∈ Γn then η(A) ∈ Γn and if A ∈ Γn∗ then η(A) ∈ Γn∗ .  By the above results we can state that, for every n, the min–max levels of the classes in ΓnV are all critical levels provided they satisfy the Palais–Smale condition, regardless any regularity property of the constraint since they are levels of the unconstrained admissible min–max classes Γn and Γn∗ . The corresponding critical points have also a characterization of unconstrained min–max points with the relative properties like estimates of the Morse index or others. Moreover, the three classes are equivalent from the point of view of the localization of the critical points near the maximum points of the terms of a minimizing sequence of sets, as stated in the next proposition.

Multiplicity techniques for problems without compactness

581

P ROPOSITION 4.10. If (An )n∈N is any minimizing sequence in ΓnV , i.e., for every n ∈ N, An ∈ ΓnV and limn (supu∈An I (u)) = cn , then the sequence of the homothetic expansions (ΣAn )n∈N is a minimizing sequence in Γn (and so, in particular, in Γn∗ ). P ROOF. The assertion easily follows by Lemma 4.5 and by Proposition 4.9.



P ROPOSITION 4.11. If (An )n∈N is any minimizing sequence in Γn∗ (and so, in particular, in Γn ) then the sequence of the traces on V(An ∩ V)n∈N is a minimizing sequence in ΓnV . P ROOF. The assertion easily follows by Lemma 4.4 and by Proposition 4.9.



4.4. Min–max classes on the double natural constraint In this subsection we shall apply the above introduced notion of genus to the study of the min–max classes on the double natural constraint. To this aim we fix π ∈ P2 , π = id, so that we simply have π(1) = 2, π(2) = 1, and consequently we fix the maps πˆ and πE as defined in the previous subsection. Let Ω ⊂ RN be given and let I : H → R be the functional defined in (1.15), related to problem (SP), with H = H01 (Ω). For every x ∈ Ω, we set, as usual, u+ (x) = max(u(x), 0), u− (x) = max(−u(x), 0). Let λ1 be the first eigenvalue of −
on Ω, we suppose λ < λ1 . Let W be the double natural constraint so defined   (4.2) W = u  u± = 0, ∇I (u) · u± = 0 ⊂ V, we set for n
1,

and

  ΓnW = A ⊂ W  A compact, γ ∗ (A)
n cn = inf sup I (u). A∈ΓnW u∈A

Since λ < λ1 then there exists a ground state level c¯ > 0 so, for every n, cn > 2c¯ since, for every u ∈ W, I (u± )
c¯ (see [14]). Let us introduce the sets


 c¯  C − = u ∈ H \ {0}  I (u)  , ∇I (u) · u
0 , 3 

  + C = u ∈ H \ {0}  I (u)  0 and so ∇I (u) · u  0 ,


 c¯  Ln = u ∈ H \ {0}  I (u)  cn − . 3

Let us observe that, being λ < λ1 , given any u ∈ H \ {0}, the function α → I (αu) grows with a positive derivative for α small as far as it reaches its maximum, then it has a negative

582

S. Solimini

derivative and tends to −∞ as α → +∞. So two constants ε, c > 0 such that εu ∈ C − and cu ∈ C + always exist. Moreover, if A ⊂ H \ {0} is any compact set, ε and c can be uniformly fixed for u ∈ A, as well as, if A ⊂ W, for u ∈ A± = {u± | u ∈ A}, since A± turn out to be also compact sets which do not contain 0. Furthermore, ε can be fixed in a maximal way and c can be fixed in a minimal way so that both the two values can be considered as functions of A. We introduce the function σA : A × Q2 → H , defined as 

 c−ε + c−ε − σA (u, α, β) = ε + (α + 1) u − ε + (β + 1) u 2 2

and the sets CA = σA (A × ∂Q2 )

and ΣA = σA (A × Q2 ).

We shall refer to ΣA as to the double homothetic expansion of A. Let us notice that if A ∈ ΓnW and supA I < cn + 3c¯ , then σA belongs (see Lemma 4.9) to the class of functions defined by   Fn = σ : A × Q2 → H  A ∈ ΓnW and 1, 2, 3 hold ,

where (1) σ (πH (x)) = −σ (x) ∀x ∈ A × Q2 , (2) σ (u, α, β) ∈ Ln and σ (u, α, β)+ ∈ C ± if α = ±1, (3) σ (u, α, β) ∈ Ln and σ (u, α, β)− ∈ C ± if β = ±1. Therefore ΣA belongs to the class of sets defined by  Γn = X ⊂ H \ {0} | ∃A ∈ ΓnW ,

∃σ : A × Q2 → H, σ ∈ Fn , s.t. X = σ (A × Q2 ) .

Let us introduce two further classes of continuous functions and sets, namely  F ∗ = ϕ : H → R2 | 1∗ , 2∗ , 3∗ hold ,

where (1*) ϕ(−u) = π(ϕ(u)), ˆ (2*) ±ϕ1 (u)
0 if u ∈ Ln and u+ ∈ C ± , (3*) ±ϕ2 (u)
0 if u ∈ Ln and u− ∈ C ± , and  Γn∗ = X ⊂ H \ {0} | X compact, L EMMA 4.7. Γn ⊂ Γn∗ .

 X = −X and ∀ϕ ∈ Fn∗ : γ ∗ X ∩ ϕ −1 (0)
n .

Multiplicity techniques for problems without compactness

583

P ROOF. Let ϕ ∈ F ∗ , A ∈ ΓnW , σ : A × Q2 → H , σ ∈ Fn be fixed. By virtue of Theorem 4.8, we have γ ∗ (σ (A × Q2 ) ∩ ϕ −1 (0))
n. By the arbitrariness of ϕ ∈ F ∗ , we get  σ (A × Q2 ) ∈ Γn∗ . L EMMA 4.8. For every A ∈ Γn∗ , we have A ∩ W ∈ ΓnW . P ROOF. Let g(u) = ∇I (u) · u cut off by a positive constant near 0 so that g(0) > 0. To prove the statement it suffices to note that the function ϕ± : u → −g(u± ) ∈ R2 belongs  to F ∗ and hence by the definition of Γn∗ the thesis follows. L EMMA 4.9. If A ∈ ΓnW then supΣA I = supA I and supCA I  supA I − 32 c. ¯ P ROOF. Of course, A ⊂ ΣA so supA I  supΣA I . Conversely, if u ∈ ΣA , u has the form u = µu¯ + + ν u¯ − with u¯ ∈ A. Since u¯ ± ∈ V, we have



 I (u) = I µu¯ + + I (ν u¯ − )  I u¯ + + I (u¯ − ) = I (u) ¯  sup I. A

If u ∈ CA , then {µ, ν} ∩ {ε, c} = ∅ so I (u¯ ± ) − I (u± )
c¯ − and “−” sign.

c¯ 3

for at least one of the “+” 

L EMMA 4.10. The min–max levels of the above defined classes are the same, i.e., inf sup I (u) = inf ∗ sup I (u) = cn .

A∈Γn u∈A

A∈Γn u∈A

P ROOF. By virtue of the inclusion Γn ⊂ Γn∗ , we have inf sup I (u)  inf sup I (u).

A∈Γn∗ u∈A

A∈Γn u∈A

Moreover, for every X ∈ Γn∗ , from Lemma 4.8 we have cn  sup  sup I (u) u∈X∩W

u∈X

and so, by the arbitrariness of X, cn  infA∈Γn∗ supu∈A I (u). Finally, by Lemma 4.9 we have that, for every X ∈ ΓnW , since ΣX ∈ Γn , inf sup I (u)  sup I (u) = sup I (u).

A∈Γn u∈A

u∈ΣX

u∈X

By the arbitrariness of X in ΓnW , we have inf sup I (u)  inf sup I (u) = cn .

A∈Γn u∈A

A∈ΓnW u∈A

P ROPOSITION 4.12. Γn and Γn∗ are two admissible min–max classes.



584

S. Solimini

P ROOF. Since cn > 0, we can find a cut-off function ϕ such that ϕ(cn ) = 1 and ϕ(s) = 0 for s  cn − 3c¯ (see [44,46]). By multiplying the gradient ∇I (u) by ϕ(I (u)) we obtain a cutoff gradient flow η : R × H → H which lets the points at level cn move along the reverse gradient direction and leaves the points in Ln fixed. It follows that, setting ηt : x → η(t, x), if σ ∈ Fn and ϕ ∈ Fn∗ then ηt ◦ σ ∈ Fn and ϕ ◦ ηt ∈ Fn∗ . So if A ∈ Γn then ηt (A) ∈ Γn and if A ∈ Γn∗ then ηt (A) ∈ Γn∗ .  By the above results we can state that, for every n, the min–max levels of the classes in ΓnW are all critical levels provided they satisfy the Palais–Smale condition, regardless any regularity property of the constraint since they are levels of the unconstrained admissible min–max classes Γn and Γn∗ . The corresponding critical points also have a characterization of unconstrained min–max points with the relative properties like, for instance, the direct estimates of the Morse index in the whole space. Moreover, the three classes are also equivalent from the point of view of the localization of the critical points near the maximum points of the terms of a minimizing sequence of sets, as stated in the next proposition. P ROPOSITION 4.13. If (Ai )i∈N is any minimizing sequence in ΓnW , i.e., for every i ∈ N, Ai ∈ ΓnW and limi (supu∈Ai I (u)) = cn , then the sequence of the double homothetic expansions (ΣAi )i∈N is a minimizing sequence in Γn (and so, in particular, in Γn∗ ) for i large. P ROOF. The assertion easily follows from Lemmas 4.9 and 4.10.



P ROPOSITION 4.14. If (Ai )i∈N is any minimizing sequence in Γn∗ (and so, in particular, in Γn ) then the sequence of the traces on W(Ai ∩ W)i∈N is a minimizing sequence in ΓnW . P ROOF. The assertion easily follows from Lemmas 4.8 and 4.10.



It is worth to remark that the two previous propositions imply, in particular, that the admissible min–max classes Γn and Γn∗ produce Palais–Smale sequences in W. 4.5. An estimate on the Morse index As an example of the utility of working without constraints and in view of an application in remaining part of this subsection, we shall briefly show a rough estimate which states that at the level cn we can find critical points in which the augmented Morse index is greater or equal to n − 1 (the sharp estimate would give: greater or equal to n). The proofs follow the ideas in [27], to which we refer for more details. Firstly we shall take into account the case in which at level cn we only have nondegenerate critical points. In such a case we shall show that each one of such points can be avoided, in the sense that every minimizing sequence can be modified in order to have an empty intersection with a neighborhood of such a point, if it has a Morse index smaller than n − 1. If all the points are in this situation, we can find a minimizing sequence of sets far away from all the critical points at level cn , getting in contradiction in case of compactness. This argument is based on a simple topological lemma, see [27].

Multiplicity techniques for problems without compactness

585

L EMMA 4.11. Let C ⊂ RN −1 be a compact set and let f : C → RN \ {0} be a continuous map. Then f has a continuous extension from RN −1 to RN \ {0}. P ROOF. By the Dugundji theorem [22] f has a continuous extension f¯ from RN −1 to RN which is C 1 out of C. By the Sard theorem f¯(RN −1 \ C) is a negligible set because it cannot contain regular values. By compactness we know that a positive number r > 0 such that Br (0) ∩ f (C) = ∅ can be chosen. Then



 Br (0) ⊂ f (C) ∪ f¯ RN −1 \ C = f¯ RN −1 .

So we can find y¯ ∈ Br (0) \ f¯(RN −1 ). We denote by p the projection of Br (0) from y¯ on ∂Br (0) extended by the identity out of Br (0). Then p ◦ f¯ is the desired extension.  Now, let u¯ be a regular critical point at level cn and let E− and E+ , respectively, denote the subspaces of H01 spanned by the eigenvectors corresponding to the negative and positive eigenvalues of ∇ 2 I (u), ¯ so that we can find a constant m > 0 such that ±∇ 2 I (u)(u ¯ ± ) · u±
mu± 2 .

∀u± ∈ E± :

(4.3)

For two given r± > 0, let B± = Br± (0) ∩ E± , S± = ∂B± in E± , B ′ = u¯ + B− + B+ and B = u¯ + (B 1 r− (0) ∩ E− ) + B+ . We shall split u ∈ B ′ as u¯ + u+ + u− with u± ∈ B± . Let 2 us divide B ′ in two parts,  B1 = u ∈ B ′ | ∇ 2 I (u)(u ¯ + ) · u+
−2∇ 2 I (u)(u ¯ − ) · u− , B2 = B ′ \ B1 .

Note that by (4.3), m u+ 2
const · u2 , 2

∀u± ∈ B1 :

±∇ 2 I (u)(u) ¯ · u+


∀u± ∈ B2 :

±∇ 2 I (u)(u) ¯ · u−  −mu− 2  −const · u2 .

(4.4) (4.5)

So, if we fix r± conveniently small, we can respectively deduce that ∀u± ∈ B1 : ∀u± ∈ B2 :

±∇I (u) · u+
0,

±I (u)  I (u) ¯ = cn .

(4.6) (4.7)

Furthermore, by also choosing r− suitably smaller than r+ we can also assume that inf

u+B ¯ − +S+

I
sn > cn .

(4.8)

In a more regular context the following lemma can be proved in a slightly easier way by using Morse lemma instead of the previous construction.

586

S. Solimini

L EMMA 4.12. Let u¯ be a nondegenerate critical point of I at level cn . Then for every A ∈ Γn∗ such that supA I < sn there exists A′ ∈ Γn∗ such that 



A′ \ B ′ ∪ −B ′ = A \ B ′ ∪ −B ′ ,

 ⊂ u¯ + E, A′ ∩ B

(4.10)

sup I  sup I.

(4.11)

A

(4.9)

A′

P ROOF. Let η : B ′ → B ′ be defined by η(u) = η(u¯ + u− + u+ ) = u¯ + u− + α(u)u+ ,

(4.12)

where α(u) = ( r2− u−  − 1)+ . So η = id on u¯ + S− + B+ . We define η in a symmetric way near the symmetric critical point −u¯ and we extend it by identity on all of H01 . We can assume that B ′ does not touch C± , so η leaves the two sets fixed. The map η so extended turns out to be discontinuous only on the sets ±(u¯ + B− + S+ ) where the functional I takes values greater or equal to sn . So A ∩ ((u¯ + B− + S+ ) ∪ −(u¯ + B− + S+ )) = ∅ and we can modify η in order to restore the discontinuity without modifying the values on A. Let A′ = η(A). Since η is odd and leaves the sets C± fixed, η(A) ∈ Γn∗ . Moreover, (4.9) follows  then by easy considerations since η = id on A \ (B ′ ∪ −B ′ ), (4.10) holds because if u ∈ B α(u) = 0 and η(u) = u¯ + u− ∈ u¯ + E− . Finally, (4.11) trivially follows from ∀u ∈ B ′ :



 I η(u)  max I (u), cn ,

(4.13)

as we are going to check. Indeed by (4.6),

 I (u) − I η(u) =



1

α(u)

∇I (u¯ + u− + su+ ) · u+ ds
0,

if u¯ + u− + α(u)u+ = η(u) is in B1 . Otherwise by (4.7), I (η(u))  cn .



The above result can be improved by the use of Lemma 4.11, if we assume that the Morse index of u, ¯ namely dim E− , is smaller than n − 1. P ROPOSITION 4.15. Let u¯ be a nondegenerate critical point at level cn and let dim E− < n − 1. Then we can find two arbitrarily small neighborhoods B and B ′ of u¯ such that (4.9) and (4.11) and A′ ∩ B = ∅

(4.14)

hold. P ROOF. Let A′ , B, B ′ be as in the previous lemma. We shall show that A′ \ (B ∪ −B) still belongs to Γn∗ and therefore it satisfies the properties in the thesis. Indeed, assume to have

Multiplicity techniques for problems without compactness

587

ϕ ∈ F ∗ and ψ ∈ Λn−1 ((A′ \ (B ∪ −B)) ∩ ϕ −1 (0)) such that 0 ∈ / ψ(((A′ \ (B ∪ −B)) ∩ −1 ϕ (0)) × Qk ). By (4.10),

   ∩ ϕ −1 (0) ⊂ u¯ + E. T = A′ \ (B ∪ −B) ∩ A′ ∩ B

We have ψ : ((A′ \ (B ∪ −B) ∩ ϕ −1 (0)) × Qk → Rn−1+k \ {0}. We want to extend ψ to ((A′ ∩ B) ∩ ϕ −1 (0)) × Qk and we must extend it from T × Qk which is the common boundary of the set on which ψ is already defined and the set to which it must be extended. We can firstly extend, by the Dugundji theorem, the components ±ψi , i = 1, . . . , k, to  ∩ ϕ −1 (0)) × F ± to positive values and then all the other components in order to ((A′ ∩ B) i have ψ : (T × Qk ) ∪ Since (T × Qk ) ∪



′   ∩ ϕ −1 (0) × ∂Qk → Rn−1+k \ {0}. A ∩B



′   ∩ ϕ −1 (0) × ∂Qk ⊂ (u¯ + E− ) × Rk A ∩B

and (u¯ + E− ) × Rk is a linear space of dimension strictly smaller than n − 1 + k, by Lemma 4.11 we can extend ψ to (T × Qk ) ∪ ((A′ ∩ B ∩ ϕ −1 (0)) × Qk ) with values in Rn−1+k \ {0}. Symmetrically, we can also extend ψ to ((A′ ∩ −B ∩ ϕ −1 (0)) × Qk ) and therefore to all of (A′ ∩ ϕ −1 (0)) × Qk with values in Rn−1+k \ {0}, in contradiction to the condition A ∈ Γn∗ .  By iterating this flattening and excision procedure, if I has only (a finite number of ) nondegenerate critical points at level cn and all of them have the Morse index strictly smaller than n − 1, we can easily find a minimizing sequence (Ai )i∈N in Γn∗ whose terms have an empty intersection with a fixed neighborhood V of the set Kcn of the critical points at level cn , getting in contradiction in case of compactness. In the general case, one can take advantage of the Marino–Prodi perturbation argument, (see [27,32,40] and [41], Chapter 3, Section 5) which allows one to reduce Kcn to a set of nondegenerate critical points and to consequently deduce the following result P ROPOSITION 4.16. If I satisfies PS at level cn there exists u¯ in Kcn which has an augmented Morse index greater or equal to n − 1. 4.6. Multiple solutions to the problem at critical growth We can now give the proof of Theorem 1.2. Let us choose a sequence (pn )n∈N in ]2, 2∗ [ such that pn → 2∗ and introduce the functionals    λ 1 1 2 2 n Iλ (v) = |∇v| − |v| − |v|pn , 2 Ω 2 Ω pn Ω whose critical points are solutions to problems (SP) for p = pn .

588

S. Solimini

Let V be as defined in Section 1.4 and Vn the analogous constraint for I = Iλn . Let ck = inf sup Iλ (v) A∈ΓkV v∈A

and ckn = inf sup Ikn (v). A∈ΓkVn v∈A

By the results of Section 4.3, we see that every ck can be also regarded as the min–max level of I on the class of sets Γk or Γk∗ . The same conclusion holds for ckn and, by a suitable choice of the sets C − and C + which can be uniformly fixed with respect to n, ckn can be seen as the min–max level Iλn on the same classes Γk and Γk∗ as before which, therefore, do not depend on n. L EMMA 4.13. If λ < λk then ck > 0 and, for every n ∈ N, ckn > 0. ⊥ = ∅. Let u ∈ A ∩ E ⊥ ⊂ V ∩ E ⊥ . P ROOF. Let A ∈ ΓkV , by Proposition 4.2 A ∩ Ek−1 k−1 k−1 By the constraint equation,     λ 1 1 1− Iλ (u) = |∇u|2 − λ |u|2
|∇u|2 . (4.15) N Ω N λk Ω Ω

The constraint equation also implies      λ ∗ |∇u|2 − λ |u|2
1 − |u|2 = |∇u|2 , λk Ω Ω Ω Ω which implies 

2∗ −2 λ S u2∗
1− λk

and therefore

sup I
I (u) = A

1 N



Ω

∗

|u|2


1 N



1−

N/2 λ S . λk

Taking the infimum for A ∈ ΓkV we finally have 1 ck
N



N/2 λ 1− S > 0. λk

The second part of the thesis can be proved in the same way. The proof of Theorem 1.2 will follow from the following two lemmas.



Multiplicity techniques for problems without compactness

589

L EMMA 4.14. limn→+∞ ckn = ck for any k as in the previous lemma. P ROOF. Let us fix k ∈ N and A ∈ Γk , then for any u ∈ A, Iλn (u) → Iλ (u). Being A compact and the functionals equicontinuous, sup Iλn (u) → sup Iλ (u).

u∈A

u∈A

Then lim supn∈N ckn  supu∈A Iλ (u) and, being A an arbitrary set in Γk , we get lim sup ckn  ck . n→+∞

∗

Since for s > 0 the function h(s) = p1n s pn − 21∗ s 2 gets its maximum value in s = 1, we have h(s)  p1n − 21∗ for all s > 0. Therefore for every u ∈ H01 , Iλ (u)  Iλn (u) +



1 1 |Ω| − pn 2 ∗

so, for any k ∈ N, ck  lim inf ckn n→+∞

and we have the thesis.



L EMMA 4.15. limk→+∞ ck = +∞. P ROOF. We want to remark that such a result is not based on compactness properties because, if we prove the statement for λ > 0, i.e., when we have compactness, the statement itself is obviously true when λ  0 and compactness fails. Moreover, this lemma is also true in lower dimension, as we can see, in the same way, by adding a suitably big subcritical term. On the other side, the use of compactness techniques takes advantage of the previous results in this chapter. Let us suppose, by contradiction, that the sequence (ck )k∈N is bounded, hence it converges to a real number c. For any k ∈ N, by Lemma 4.14 there exists nk > k such that |cknk − ck | < k1 ; hence lim cnk k→+∞ k

= lim ck = c k→+∞

(4.16)

and the sequence (nk )k∈N is diverging, i.e., lim nk = +∞.

k→+∞

Let unk be a solution of (SP) at level cknk . Using the Morse index estimates on min–max points as in Section 4.5, we can select the sequence (unk )k∈N such that every unk has an

590

S. Solimini

augmented Morse index greater or equal to k − 1. By our assumptions, we can claim that the sequence (unk )k∈N is bounded in H01 (Ω). Indeed, being unk a solution to (SP) we have Iλnk (unk ) =



1 1 − 2 pnk



Ω

|unk |pnk → c,

(4.17)

which gives the boundedness of −
unk in H −1 and, in turn, the boundedness of unk on H01 . So the sequence (unk )k∈N is uniformly bounded by Theorem 1.1 and therefore the Morse index of unk must keep bounded, in contradiction to our construction.  P ROOF OF T HEOREM 1.2. Fixed k ∈ N, we take, for any n ∈ N, un = unk a critical point at level ckn for the functional Iλn . First we use Lemma 4.14 and the analogous of (4.17) to have (un )n∈N bounded in H01 . The sequence (un )n∈N , by Theorem 1.1, is then uniformly bounded, so by standard compactness arguments we can find a converging subsequence to a solution u¯ k to (CP) at level ck , as follows from Lemma 4.14. By Lemma 4.15, we have infinitely many distinct values of ck for k ∈ N and so the proof is concluded.  4.7. Multiple solutions to the problem on the whole domain This part is devoted to the proof of Theorem 1.4 which states the existence of infinitely many solutions to problem (P). Let us fix a sequence (rn )n∈N , rn ∈ R+ such that rn → +∞ and consider the problems (Pn ) = (APrn ) defined in Section 1.2. Beside the functional I defined in Section 1.2, we introduce the functional  

 1 1 In (u) = |∇u|2 + a(x)u2 dx − |u|p dx, 2 Brn (0) p Brn (0) and the min–max levels ck = inf sup I (v) A∈ΓkV v∈A

and ckn = inf sup In (v), A∈ΓkVn v∈A

where V is the natural constraint related to the functional I and Vn is the natural constraint related to In or, equivalently, Vn = V ∩ H01 (Brn (0)). Since (a1 ) and (a2 ) widely imply that (a − 21 a∞ )− = sup(−(a − 21 a∞ ), 0) ∈ LN/2 , we know that the eigenvalue problem 

− −
u = µ a − 12 a∞ (x)u in RN , (LP)

 u ∈ H 1 RN ,

Multiplicity techniques for problems without compactness

591

has a diverging sequence of eigenvalues µ1 < µ2  µ3  · · ·  µn  · · · , see [30]. So we can find k such that µk > 1. Also in this situation we have the analogous statement of Lemma 4.13. L EMMA 4.16. If µk > 1 then ck > 0. P ROOF. The proof is formally equal to that of Lemma 4.13, provided we take the functions ϕi as the eigenvectors given by (LP) corresponding to the eigenvalues µi and ⊥ , Ei = Span{ϕ1 , ϕ2 , . . . , ϕi }. By the constraint equation, if u ∈ V ∩ Ek−1 

I (u) =

1 1 − 2 p



RN

|∇u|2 + a(x)u2



−  1 1 |∇u|2 + a∞ u2 − a − a∞ (x)u2 2 2 RN     1 1 a∞ 2 1
|∇u|2 + 1− − u . 2 p µk RN RN 2 




1 1 − 2 p





(4.18)

So    1 a∞ 2 1− |∇u| + u2 u
µk 2 RN RN   1 a∞ 1 1 2 2 1− Su2∗ + − u2 .
2 p µk 2 

p

1 1 − 2 p

(4.19)

On the other hand, by the Hölder and Young inequalities, 2(2∗ (p−2))/(p(2∗ −2))

u2p  u2∗ 

2(2(2∗ −p))/(p(2∗ −2))

u2

2∗ (p − 2) 2(2∗ − p) u22∗ + u22 . ∗ p(2 − 2) p(2∗ − 2)

(4.20)

Therefore, combining (4.19) and (4.20), one easily gets p

up 2  const · up and finally, I (u) =



1 1 − 2 p



RN

|u|p
const.

⊥ = ∅, by fixing u ∈ Since, for every A ∈ ΓkV , one has by Proposition 4.2 that A ∩ Ek−1 ⊥ ⊂ V ∩ E ⊥ , we have A ∩ Ek−1 k−1

sup I
I (u)
const A

(4.21)

592

S. Solimini

and, taking the infimum, ck
const > 0.



In this second case we always have ck  ckn

(4.22)

because Vn = V ∩ H01 (Brn ) ⊂ V. Moreover, the property of Lemma 4.13 still holds. L EMMA 4.17. If µk > 1 then ck = limn ckn . P ROOF. Let k be given and let A ∈ ΓkVn . Since A is a compact set, for every ε > 0 A has a finite ε-net F contained in D(Ω). For n large, F ⊂ H01 (Brn ). Let Pn be the orthogonal projection of H 1 (RN ) onto H01 (Brn ). We know that, for every x ∈ A, x − Pn (x) < ε. Since A ⊂ V, for v ∈ Pn (A),   1/(p−2) 2− 2 RN |∇v| RN a(x)v dx  α(v) = > 0, (4.23) p RN |v| dx provided ε is taken sufficiently small. Then the map ηn : A → H10 (Brn ) ∩ V = Vn defined by

 ηn : u → α Pn (u) Pn (u) (4.24)

is an odd map. So γ ∗ (ηn (A))
γ ∗ (A)
k and therefore, ηn (A) ∈ ΓkVn . One can also easily see that for any δ > 0, by taking a conveniently small value of ε to have ηn (A) close enough to A, ckn  sup I  sup I + δ ηn (A)

(4.25)

A

for n large. Taking the infimum for A ∈ ΓkV and δ > 0 we finally find lim sup ckn  ck ,

(4.26)

n

which, combined with (4.22), gives the thesis.



We can also formally use the same proof of Lemma 4.13 to see that, also in this second case, the sequence ck is diverging. The same argument used for the proof of Theorem 1.2 gives now Theorem 1.4. P ROOF OF T HEOREM 1.4. Let k be as in Lemma 4.16. For every n we can find a solution un to (Pn ) at level ckn . Since ckn

= In (un ) =



1 1 p un p → ck − 2 p

Multiplicity techniques for problems without compactness

593

as n → +∞, then (un )n∈N is bounded in Lp , so also   |∇un |2 + a(x)u2n  const RN

RN

and finally, 

1 |∇un |2 + a∞ 2 RN



RN

− 1 a − a∞ u2n 2 RN  − (p/2)′   1  un 2p  const +  a − a∞   2

u2n  const +





 const,

so (un )n∈N is bounded in H 1 (RN ). Then by Theorem 1.3, (un )n∈N has a converging subsequence to a solution uk to (P) at level ck . Since (ck )k∈N is diverging, (P) has infinitely many solutions. 

5. Concluding remarks We conclude this exposition with some comments on the assumptions employed for the two problems and indicating what one knows or what one can expect without them. In the second part of the section we shall show what kind of multiplicity results are already known for (SP) in nonsymmetric cases when the fundamental restriction N
7 on the dimension is not assumed.

5.1. Review on the assumptions and open problems We have considered the critical growth problem only from the point of view of taking into exam the presence of subcritical terms which make the concentrations to be not convenient. The techniques discussed here have no application (so far) to problems in which the concentrations are avoided thanks to assumptions on the shape of the domain or other geometric conditions. Easy extensions are possible to similar equations involving other operators as, for instance, the p-Laplacian or more general elliptic operators. The most delicate point is concerned with the bound on the dimension. The dimension seven has no particular meaning as happens, for instance, with the minimal surface problems, but depends on the linearity of the subcritical perturbation term λu. A nonlinear subcritical term would require a different bound on the dimension. Let us point out that N
4 is enough for the existence of a nontrivial solution for every λ but three more dimensions are needed for the multiplicity results. The reason is due to the fact that as we have seen in Section 1.5, starting on dimension 4, the gain due to the quadratic term −λ Ω u2 in the functional is more consistent than the cost of a cut-off term which brings a Talenti function to zero within the boundary of Ω. On the other hand, if one wants to perform a similar test for the

594

S. Solimini

multiplicity problem, one must cut off the function in order to reach a value of opposite sign and so the cost of the cut-off is considerably more expensive. As we have seen in Section 2, the existence of infinitely many solutions to (CP) is assured for every Ω ⊂ RN only for N
7 and relies on a compactness theorem which cannot be extended to lower dimensions. This circumstance does not mean that the existence of infinitely many solutions cannot be proved through different tools as it happens, for instance, in the case of symmetric domains (see [24]). Such a result is false if one looks for particular solutions as, if Ω is a ball, the radial ones (see [2]), which certainly exist if N
7, and this suggests, in the general case, to search for solutions which change sign near the boundary, so excluding the radial symmetric functions. This approach was pursued in [14] in order to show the existence of two (pairs of ) distinct solutions for N
4 and λ < λ1 . A result in this sense will be shown in the second part of this section. In the case of problem (P) the assumptions (a1 ) and (a2 ) let the superquadratic term of the functional bounded in terms of the quadratic part and allow the proof of Proposition 1.1. The smoothness asked in (a1 ) is more than what we need to this aim (a − ∈ LN/2 (RN ) would be more than enough) but it is required for the other assumptions. Condition (a3 ) is the equivalent of the bound N
7 for (CP). The proof of the multiplicity results without this assumption would probably require completely different arguments. One can possibly show even uniqueness results in some particular cases, as happens for (CP) with the radial symmetry. On the contrary (a4 ) has been only used for the proof of Lemma 2.4 to the to an integral of ∂∂ax4 . It can certainly be aim of passing from an integral of the function ∂a ∂ t4 weakened, the question if some assumption of this kind is necessary for the multiplicity result or what is the minimal condition is open.

5.2. Estimates in lower dimension Now we shall prove, following [20], that the estimate in [14] increases up to at least N2 + 1 solutions or even to N + 1 if λ is suitably close to zero. This result has been recently extended in [15] to the case λ
λ1 . In any case, there is no reason to suppose that such a result may be optimal and the problem of proving the existence of infinitely many solutions, or even of getting an optimal estimate on the number of solutions, remain, as far as we know, largely open for N  6. We shall work with the double natural constraint introduced in Section 4.4. L EMMA 5.1. Let (un )n∈N be a PS sequence in W at level c < cannot converge weakly to zero.

2 N/2 , NS

then the sequence

P ROOF. Since (un )n∈N is a PS sequence in W and λ ∈ ]0, λ1 [, we have ∇u± n 
const therefore an eventual strong limit of the two sequences (u± n )n∈N cannot be zero. We have to extend this claim to the case of a weak limit, which brings a weaker information when the sequence is not compact. We know that the “bad” levels c for noncompact PS sequences detected by Theorem 1.6 in ]0, N2 S N/2 [ are c = c′ + N1 S N/2 with c′ critical level for Iλ and the weak limit ϕ0 of the sequence is a solution u to (P) at level c′ . The obstruction to the compactness of (un )n∈N is given by scaled copies of a global solution ϕ1 which has a

Multiplicity techniques for problems without compactness

595

constant sign and disappears if we restrict ourselves to u± n for one of the + or − signs. So, ± strongly and so u = 0, according to the assertion in one of the two cases, we have u± → u n in the beginning of the proof. Therefore or c or c − N1 S N/2 is a nonnull critical level which corresponds to the strong or respectively to the weak limit of the sequence (un )n∈N .  We set c0 = inf Iλ (u). u∈V

L EMMA 5.2. For all k ∈ {1, . . . , N + 1}, we have 2c0  ck
0 contained in Ω which, in order to use a simpler notation, will be assumed, without any restriction, to be centered in the origin. For any ν ∈ S N −1 = ∂B, we consider two balls B1ν and B2ν of radius R2 contained in B and tangent to ∂B respectively in ν and in −ν. Let us consider the function ϕ : ∂B → W such that for any ν ∈ ∂B, ϕ(ν) = uν : Ω → R, where (uν )+ and (uν )− are respectively the function u∗σ found in Section 1.5 in B1ν and B2ν and let us call c¯ = Iλ ((uν )+ ) = Iλ ((uν )− ). In this way, we map ∂B into a set A ⊂ W in an odd continuous way, therefore γ (A)
γ (∂B) = N . Moreover, ∀ν ∈ ∂B: Iλ (uν ) = Iλ ((uν )+ ) + Iλ ((uν )− ) = 2c¯ < N2 S N/2 . From the bound cN < 2c¯ we shall now pass to show that cN +1 is below N2 S N/2 . Actually we shall find in this new case the bound cN +1  c¯ + N1 S N/2 , by extending the above introduced map ϕ from ∂B to an N -dimensional sphere. From the construction in Section 1.5, one can easily see that u¯ 1 can be modified continuously to the analogous function u¯ s1 , which has a support on the ball B ν (s) concentric with B1ν and radius s, belongs to V and keeps the property Iλ (u¯ s1 ) < N1 S N/2 . Of course Iλ (u¯ s1 ) → N1 S N/2 as s → 0. We shall firstly extend ϕ to B, namely we shall find a continuous homotopy with values in W from ϕ to a constant map. This homotopy will be performed in two steps. Firstly we shall shrink the radius of B1ν to a conveniently small radius ρ by taking H1 (s, ν) = u¯ s1 − u¯ 2 with s ∈ [0, R2 ]. Then we shall translate the balls B ν (ρ) and B2ν bringing the two centers on the origin, obtaining the homotopy H2 (t, ν) = x

ρ  u¯ 1 →

  ν ν − u¯ 2 x − t x +t 2 2

596

S. Solimini

for t ∈ [0, 1]. ρ The very different scales of u¯ 1 and u¯ 2 make infinitesimal the mixed terms in the evaluation of Iλ , so we have, for all t ∈ [0, 1],

ρ  1 Iλ H2 (t, ν) = Iλ u¯ 1 + Iλ (u¯ 2 ) + ε1 (ρ) = c¯ + S N/2 + ε2 (ρ), N

(5.1)

where for i = 1, 2, εi (ρ) → 0 when ρ → 0. The definition of H2 should still be changed by multiplying the positive and the negative part of H2 (t, ν) by coefficients α± (t) in such a way to provide that H2 (t, ν) ∈ W for all t ∈ [0, 1]. Since the function α± (t) are clearly continuous and α± (t) → 1 as ρ → 0, as we can see by arguing as in (5.1), we still keep (5.1) after this small correction. H2 (1, ν) gives the same function whatever is ν ∈ ∂B, so the homotopy connects ϕ to a constant map. We have in this way an extension of ϕ to B and so to a hemisphere in dimension N + 1. By an odd extension we define ϕ on the whole sphere. Then the set of the functions obtained via this transformation is a compact symmetric set contained in W whose genus is greater or equal to N + 1, moreover, on this set the functional Iλ is bounded by c¯ + N1 S N/2 + ε2 (ρ) < N2 S N/2 and the thesis follows.  R EMARK 5.1. If we work on the larger constraint V, we can add one more step to the previous construction showing that even the constrained level cN +2 is in such a case lower than N2 S N/2 . Indeed, one can easily find a homotopy from the map ϕ, extended to the sphere in N + 1 dimension to a constant map, by arguing as follows. For s ∈ [0, 1] one performs the same construction as before by multiplying by s the function u∗σ taken in B2ν and then normalizing the whole uν by multiplying it by a normalization constant α(s) which lets α(s)uν ∈ V. Analogously, one normalizes Hi (s, ν). In this way we can, roughly speaking, “kill” the negative part of the function remaining only with scaled copies of u∗σ on some balls contained in B. Then it is easy to modify such functions continuously reconducing all of them to the function u∗σ defined as in Section 1.5 on the whole ball. This homotopy extends ϕ to a hemisphere in dimension N +2. In any case, even with the present approach, we gain a further critical level by adding to the levels c1 , c2 , . . . , ck+1 detected on W the level, which we have denoted by c0 , obtained as the infimum of I on V, so that the difference apparently relies in a shift of the index which runs from 0 to N + 1 rather than from 1 to N + 2. However, the further information that, for i
1, ci is a critical level corresponding to a PS sequence in W will allow the use of Lemma 5.1 and motivates the choice of working on the double natural constraint. Nevertheless, approaching the problem on V would only give the further difficulty of avoiding the level N1 S N/2 in Lemma 5.1 and would finally lead to prove the same result as Theorem 5.1 in even dimension but to find one solution less in odd dimension. On the other hand, Theorem 5.2 can be proved also in this other way. We shall now observe that if two different levels ci coincide, then (P) must necessarily have infinitely many solutions, even if the [P.S.]ci condition fails. Note that this lack of compactness make us unable even to say that such a level must be critical, however Lemma 3.2 permits to conclude that the number of solutions at a possibly lower level is necessarily infinite.

Multiplicity techniques for problems without compactness

597

L EMMA 5.3. If there exist i, j ∈ {1, . . . , N + 1}, i = j , such that ci = cj , then (P) has uncountably many solutions. P ROOF. We can reduce ourselves to the case in which j = i + 1 with i ∈ {1, . . . , N }. We shall prove that, in this hypothesis, we can find a solution to (P) which is orthogonal N be a minimizing sequence, i.e., to every given function v ∈ H01 (Ω). Let (An )n∈N ∈ Γi+1 1 lim supn (supAn Iλ ) = ci+1 . Let us fix v ∈ H0 (Ω) and consider for every n ∈ N A′n = An ∩ v ⊥ . The sequence (A′n )n∈N is, by Lemma 4.3, a sequence in Γi and, being 4 3 4 3 lim sup sup Iλ  lim sup sup Iλ = ci+1 = ci , n

A′n

n

An

it is a minimizing sequence. Let (un )n∈N be a constrained PS sequence at level ci close to the sequence (A′n )n∈N (i.e., limn d(un , An ) = 0), then by Lemma 5.1 its weak limit u¯ is a nontrivial solution to (CP) which is orthogonal to v. If (CP) has a only countably many solutions, the set of the functions which are orthogonal to some solution to (CP) is a first category set. So we can find a function v ∈ H01 (Ω) whose scalar product with every nontrivial solution is not null and we find a contradiction.  T HEOREM 5.1. Let N
4, then (CP) has at least ∀λ ∈ ]0, λ1 [.

N 2

+ 1 distinct ( pairs of ) solutions

P ROOF. Taking into account Lemma 5.3, we can suppose that ∀i, j ∈ {1, . . . , N + 1}, ci = cj so that we have N + 1 distinct levels ci for i = 1, . . . , k + 1. Adding the ground state level c0 , obtained as minimal level on V, we have N + 2 distinct levels 0 < c0 < c1 < · · · < cN +1
N1 S N/2 . This last property implies that ∀i ∈ {2, . . . , N + 1}, ci − N1 S N/2 < c1 , so the only critical value which can be given by two different min–max  approaches of the type considered above is c0 .

598

S. Solimini

Acknowledgment The author is grateful to Francesco Maddalena for his continuous support in organizing notes provided in the Introduction.

References [1] A. Ambrosetti, M. Badiale and S. Cingolani, Semiclassical states of nonlinear Schrödinger equations, Arch. Ration. Mech. Anal. 140 (1997), 285–300. [2] F.V. Atkinson, H. Brezis and L.A. Peletier, Solutions d’Equations elliptiques avec exposant de Sobolev critique qui changent de signe, C. R. Math. Acad. Sci. Paris 306 (1988), 711–714. [3] F.V. Atkinson, H. Brezis and L.A. Peletier, Nodal solutions of elliptic equations with critical Sobolev exponents, J. Differential Equations 85 (1990), 151–170. [4] A. Bahri and P.L. Lions, On the existence of a positive solution of semilinear elliptic equations in unbounded domains, Ann. Inst. H. Poincaré Anal. Non Linéaire 14 (1997), 365–413. [5] T. Bartsch and Z.Q. Wang, Sign changing solutions of nonlinear Schrödinger equations, Topol. Methods Nonlinear Anal. 13 (1999), 191–198. [6] T. Bartsch and M. Willem, Infinitely many nonradial solutions of an Euclidean scalar field equation, J. Funct. Anal. 117 (1993), 447–460. [7] V. Benci and G. Cerami, Positive solutions of some nonlinear elliptic problems in exterior domains, Arch. Ration. Mech. Anal. 99 (1987), 283–300. [8] H. Berestycki and P.L. Lions, Nonlinear scalar field equations, I. Existence of a ground state, II. Existence of infinitely many solutions, Arch. Ration. Mech. Anal. 82 (1983), 313–346, 347–376. [9] M.S. Berger, On the existence and structure of stationary states for a nonlinear Klein–Gordon equation, J. Funct. Anal. 9 (1972), 249–261. [10] H. Brezis and T. Kato, Remarks on the Schrödinger operator with singular complex potential, J. Math. Pures Appl. 58 (1979), 137–151. [11] H. Brezis and L. Nirenberg, Positive solutions of nonlinear elliptic equations involving critical Sobolev exponents, Comm. Pure Appl. Math. 36 (1983), 437–477. [12] A. Capozzi, D. Fortunato and G. Palmieri, An existence result for nonlinear elliptic problems involving critical Sobolev exponent, Ann. Inst. H. Poincaré Anal. Non Linéaire. 2 (1985), 463–470. [13] G. Cerami, G. Devillanova and S. Solimini, Infinitely many bound states for some nonlinear scalar field equations, Calc. Var. Partial Differential Equations 23 (2005), 139–168. [14] G. Cerami, S. Solimini and M. Struwe, Some existence results for superlinear elliptic boundary value problems involving critical exponents, J. Funct. Anal. 69 (1986), 289–306. [15] M. Clapp and T. Weth, Multiple solutions for the Brezis–Nirenberg problem, Preprint. [16] C.V. Coffman, Uniqueness of the ground state solution for −
u − u + u3 = 0 and a variational characterization of other solutions, Arch. Ration. Mech. Anal. 46 (1972), 81–95. [17] S. Coleman, V. Glaser and A. Martin, Action minima among solutions to a class of Euclidean scalar field equations, Comm. Math. Phys. 58 (1978), 211–221. [18] V. Coti Zelati and P. Rabinowitz, Homoclinic type solutions for a semilinear elliptic PDE on RN , Comm. Pure Appl. Math. 10 (1992), 1217–1269. [19] G. Devillanova and S. Solimini, Concentrations estimates and multiple solutions to elliptic problems at critical growth, Adv. Differential Equations 7 (2002), 1257–1280. [20] G. Devillanova and S. Solimini, A multiplicity result for elliptic equations at critical growth in low dimension, Comun. Contemp. Math. 5 (2003), 171–177. [21] W.-Y. Ding and W.-M. Ni, On the existence of positive entire solutions of a semilinear elliptic equation, Arch. Ration. Mech. Anal. 91 (1986), 283–308. [22] J. Dugundji, Topology, 8th Edition, Allin and Bacon, Boston, MA (1973). [23] A. Floer and A. Weinstein, Nonspreading wave packets for the cubic Schrödinger equation with a bounded potential, J. Funct. Anal. 69 (1986), 397–408.

Multiplicity techniques for problems without compactness

599

[24] D. Fortunato and E. Jannelli, Infinitely many solutions for some nonlinear elliptic problems in symmetrical domains, Proc. Roy. Soc. Edinburgh Sect. A 105 (1987), 205–213. [25] E. Jannelli and S. Solimini, Concentration estimates for critical problems, Ricerche Mat. Sect. A XLVIII (1999), Supplemento, 233–257. [26] M.A. Krasnoselskii, Topological Methods in the Theory of Nonlinear Integral Equations, Macmillan, New York (1964). [27] A. Lazer and S. Solimini, Nontrivial solutions of operator equations and Morse indices of critical points of min–max type, Nonlinear. Anal. 12 (1988), 761–775. [28] P.L. Lions, The concentration compactness principle in the calculus of variations, Parts I and II, Ann. Inst. H. Poincaré Anal. Non Linéaire 1 (1984), 109–145, 223–283. [29] P.L. Lions, The concentration–compactness principle in the calculus of variations. The limit case – Part II, Rev. Mat. Iberoamericana 12 (1985), 45–121. [30] A. Manes and A.M. Michelettii, Un’estensione della teoria variazionale classica degli autovalori per operatori ellittici del secondo ordine, Boll. Unione Mat. Ital. 8 (1973), 285–301. [31] C. Maniscalco, Multiple solutions for semilinear elliptic problems in RN , Ann. Univ. Ferrara 37 (1991), 95–110. [32] A. Marino and G. Prodi, Metodi perturbativi della teoria di Morse, Boll. Unione Mat. Ital. 11 (1975), 1–32. [33] C. Miranda, Un’osservazione sul teorema di Brower, Boll. Unione Mat. Ital. Sez. II XIX (1940), 5–7. [34] R. Molle, M. Musso and D. Passaseo, Positive solutions for a class of nonlinear elliptic problems in RN , Proc. Roy. Soc. Edinburgh Sect. A 130 (2000), 141–166. [35] Z. Nehari, On a nonlinear differential equation arising in nuclear physics, Proc. Roy. Irish Acad. 62 (1963), 117–135. [36] Y.G. Oh, On positive multi bump bound states of nonlinear Schrödinger equations under multiple wall potential, Comm. Math. Phys. 131 (1990), 223–253. [37] S. Pohozaev, Eigenfunctions of the equation
u + λf (u) = 0, Soviet Mat. Dokl. 6 (1965), 1408–1411. [38] P.H. Rabinowitz, On a class of nonlinear Schrödinger equations, Z. Angew. Math. Phys. 43 (1992), 270–291. [39] S. Solimini, On the existence of infinitely many radial solutions for some elliptic problems, Rev. Mat. Apl. 9 (1986), 75–86. [40] S. Solimini, Morse index estimates in min–max theorems, Manuscripta Math. 63 (1989), 421–453. [41] S. Solimini, Lecture notes on some estimates in functional analysis via convolution techniques, Internal Report 91, 1, I.R.M.A. – Bari, Italy (1991). [42] S. Solimini, A note on compactness-type properties with respect to Lorentz norm of bounded subsets of a Sobolev spaces, Ann. Inst. H. Poincaré Anal. Non Linéaire 12 (1995), 319–337. [43] S. Solimini, Min–Max levels on the double natural constraint, Preprint. [44] S. Solimini, Notes on Min–Max Theorems, Lecture Notes SISSA, Trieste, Italy (1989). [45] M. Struwe, A global compactness result for elliptic boundary value problems involving limiting nonlinearities, Math. Z. 187 (1984), 511–517. [46] M. Struwe, Variational Methods: Applications to Nonlinear PDE & Hamiltonian Systems, Springer-Verlag, Berlin (1990).

This page intentionally left blank

Author Index Roman numbers refer to pages on which the author (or his/her work) is mentioned. Italic numbers refer to reference pages. Numbers between brackets are the reference numbers. No distinction is made between first and co-author(s).

Abreu, E. 315, 375, 401 [1] Acerbi, E. 73, 123 [1] Adam, C. 491, 512 [1]; 512 [2] Adami, R. 421, 512 [3] Adams, D. 433, 436, 441, 442, 512 [4] Adams, R.A. 316, 402 [2] Adimurthi 344, 361, 368, 369, 402 [3] Aftalion, A. 31, 33, 52 [1] Aharonov, Y. 421, 487, 512 [5]; 512 [6] Ahmedou, M.O. 314, 344, 346, 404 [60]; 404 [61] Aizenman, M. 469, 512 [7] Akilov, G.P. 460, 514 [60] Ali, S.W. 169, 205 [AC] Alibert, J.J. 65, 123 [2] Allaire, G. 62, 123 [3] Amann, H. 129, 147, 156, 166, 205 [Am]; 226, 305, 306 [1]; 306 [2]; 306 [3]; 314, 321, 402 [4] Ambrosetti, A. 3, 5, 31, 33, 52 [2]; 52 [3]; 52 [4]; 52 [5]; 298, 306 [4]; 314, 325, 326, 344, 346, 402 [5]; 402 [6]; 526, 598 [1] Anane, A. 351, 402 [7] Andreu, F. 315, 370, 371, 396, 402 [8]; 402 [9]; 402 [10]; 402 [11]; 402 [12] Arcoya, D. 315, 321, 395, 396, 399, 400, 402 [13]; 402 [14] Arendt, W. 460, 512 [8] Aris, R. 192, 206 [Ar] Atkinson, F.V. 522–524, 594, 598 [2]; 598 [3] Aubert, G. 61, 123 [4]; 123 [5]; 124 [6] Aubin, T. 313, 369, 402 [15] Avron, J. 474, 482, 492, 494, 512 [9]

Ball, J.M. 63, 121, 124 [7]; 124 [8] Ballester, C. 371, 402 [8] Bandle, C. 218, 301, 303, 306 [5]; 306 [6]; 306 [7] Bandyopadhyay, S. 81, 124 [9] Barles, G. 313, 402 [16] Barroso, A.C. 81, 124 [9] Bartsch, T. 17, 24, 27, 29, 31, 33, 34, 38, 40–42, 51, 52 [10]; 52 [11]; 52 [12]; 52 [13]; 52 [14]; 52 [15]; 52 [16]; 52 [17]; 52 [18]; 52 [19]; 52 [20]; 387, 391, 393, 402 [17]; 525, 526, 598 [5]; 598 [6] Bauman, P. 61, 124 [10] Bebernes, J. 170, 206 [BE] Beckner, W. 344, 402 [18] Benci, V. 17, 52 [21]; 52 [22]; 52 [23]; 525, 527, 533, 598 [7] Benguria, R. 472, 512 [11] Bénilan, Ph. 315, 396, 402 [19] Berestycki, H. 18, 20, 35, 40, 52 [7]; 52 [24]; 52 [25]; 53 [26]; 525, 598 [8] Berezanski, Yu. 426, 512 [12] Berezin, F. 409, 420, 430, 446, 512 [13] Berger, M.S. 525, 598 [9] Biezuner, R.J. 369, 402 [20] Birman, M.Sh. 409, 412, 414, 445–447, 456, 464, 467, 507, 510, 511, 512 [14]; 512 [15]; 512 [16]; 512 [17]; 512 [18] Blanchard, Ph. 472, 512 [19] Blat, J. 150, 163, 169, 206 [BB1]; 206 [BB2] Bohm, D. 421, 512 [5] Bolle, Ph. 35, 53 [27]; 53 [28] Bony, J.M. 139, 206 [Bo] Borzov, V.V. 412, 512 [15] Bourgain, J. 318, 402 [21] Braverman, M. 421, 428, 431, 432, 512 [20] Bressan, A. 74, 81, 124 [11]

Badiale, M. 526, 598 [1] Bahri, A. 8, 18, 35, 52 [6]; 52 [7]; 52 [8]; 52 [9]; 525, 526, 533, 598 [4] Balinsky, A. 491, 512 [10] 601

602

Author Index

Brezis, H. 7, 8, 12, 18, 20, 21, 27, 52 [3]; 53 [29]; 53 [30]; 53 [31]; 53 [32]; 53 [33]; 297, 299, 306 [8]; 318, 330, 339, 367, 368, 402 [21]; 402 [22]; 522–524, 546, 555, 557, 594, 598 [2]; 598 [3]; 598 [10]; 598 [11] Brock, F. 9, 17, 53 [34]; 315, 395, 402 [23] Browder, F. 299, 306 [9] Brown, K.J. 150, 153, 163, 169, 170, 206 [BB1]; 206 [BB2]; 206 [BIS]; 207 [DB] Bugliaro, L. 499, 500, 502, 512 [21]; 513 [22] Bukhvalov, A.V. 460, 512 [8] Buttazzo, G. 61, 69, 124 [12]; 124 [13] Byeon, J. 12, 53 [35] Cabré, X. 314, 344, 402 [24] Calderon, A.P. 317, 402 [25] Cano-Casanova, S. 213, 299, 304, 306 [10]; 306 [11]; 306 [12]; 307 [13] Cantrell, R.S. 169, 206 [CC]; 206 [CCH] Capozzi, A. 522, 598 [12] Carriao, P. 315, 375, 401 [1] Caselles, V. 371, 402 [8] Casher, A. 487, 512 [6] Castro, A. 31, 53 [36]; 53 [37]; 170, 206 [CS] Catrina, F. 12, 53 [38] Celada, P. 61, 124 [14]; 124 [15] Cellina, A. 61, 62, 74, 81, 97, 114, 124 [16]; 124 [17]; 124 [18]; 124 [19] Cerami, G. 17, 52 [3]; 52 [21]; 52 [22]; 52 [23]; 521, 523, 525–527, 533, 581, 594, 598 [7]; 598 [13]; 598 [14] Cesari, L. 61, 124 [20]; 124 [21] Chabrowski, J. 3, 53 [39]; 53 [40]; 336, 375, 402 [26]; 402 [27] Chang, K.C. 3, 30, 53 [41] Chanillo, S. 465, 513 [23] Chavel, I. 466, 513 [24] Cherkaev, A. 365, 402 [28] Cherkaeva, E. 365, 402 [28] Chernoff, P.R. 427, 513 [25] Cherrier, P. 313, 320, 321, 356, 402 [29] Chipot, M. 314, 341–343, 403 [30]; 403 [31]; 403 [32]; 403 [33] Chlebík, M. 314, 315, 343, 368, 403 [30]; 403 [34]; 403 [35] Cingolani, S. 526, 598 [1] Cîrstea, F. C., St. ¸ 302, 303, 307 [14]; 307 [15]; 307 [16]; 307 [17]; 342, 403 [36]; 403 [37] Clapp, M. 523, 594, 598 [15] Clement, P. 395, 396, 403 [38] Cobos, F. 473, 513 [26] Coffman, C.V. 11, 12, 53 [42]; 53 [43]; 525, 598 [16]

Coleman, S. 525, 598 [17] Colombo, G. 61, 124 [18] Consul, N. 314, 403 [39] Conti, M. 27, 31, 53 [44]; 53 [45] Conway, E. 199, 206 [CHS] Coron, J.-M. 8, 18, 52 [8] Cortázar, C. 398, 403 [40] Cosner, C. 169, 205 [AC]; 206 [CC]; 206 [CCH] Cossio, J. 31, 53 [36]; 53 [37] Costa, D.G. 16, 34, 53 [46]; 53 [47] Coti Zelati, V. 36, 45–49, 53 [48]; 526, 598 [18] Courant, R. 301, 307 [18]; 409, 513 [27] Crandall, M.G. 130, 131, 177, 189, 206 [CR1]; 206 [CR2]; 237, 239, 307 [19]; 307 [20]; 315, 396, 402 [19] Cutrì, A. 61, 124 [22] Cwikel, M. 452, 455, 513 [28] Cycon, H.L. 409, 425, 492, 513 [29] Dacorogna, B. 61, 62, 64, 65, 67–69, 72–74, 76–79, 81–84, 86, 91, 92, 94, 97, 105, 109, 111, 113–117, 121–123, 123 [2]; 124 [9]; 124 [12]; 124 [23]; 124 [24]; 124 [25]; 124 [26]; 124 [27]; 124 [28]; 124 [29]; 124 [30]; 124 [31]; 124 [32]; 124 [33]; 125 [34]; 125 [35]; 125 [36]; 125 [37] Dancer, E.N. 31, 53 [49]; 53 [50]; 144, 162, 165, 168–171, 189, 191–193, 200, 205, 206 [DD1]; 206 [DD2]; 206 [DLO]; 206 [Da1]; 206 [Da2]; 206 [Da3]; 206 [Da4]; 206 [Da5]; 206 [Da6]; 206 [Da7]; 206 [Da8]; 299, 304, 307 [21]; 307 [22] Davila, J. 315, 344, 395, 397, 398, 403 [41]; 403 [42] De Blasi, F.S. 74, 81, 125 [38] de Figueiredo, D.G. 21, 53 [51]; 314, 378, 379, 385–387, 391, 393, 402 [17]; 403 [43]; 403 [44]; 403 [45] De la Bretéche, R. 470, 513 [30] de Mottoni, P. 205, 206 [dMR] Deimling, K. 129, 156, 206 [De] del Pino, M. 144, 206 [dP]; 298, 307 [26]; 315, 347, 356, 360, 361, 403 [46] Delgado, M. 303, 304, 307 [23]; 307 [24]; 307 [25] Demengel, F. 315, 403 [47] Denzler, J. 364–366, 403 [48] Devillanova, G. 521, 523, 524, 526, 532, 594, 598 [13]; 598 [19]; 598 [20] Ding, W.-Y. 526, 598 [21] Ding, Y. 51, 52 [11] Drábek, P. 358, 399, 403 [49] Druet, O. 313, 403 [50] Du, Y. 31, 53 [49]; 53 [50]; 133, 137, 142, 144, 147, 150, 153, 158, 163, 164, 167–171, 189,

Author Index 192, 200, 206 [DD1]; 206 [DD2]; 206 [Du1]; 206 [Du2]; 207 [DB]; 207 [DG]; 207 [DH]; 207 [DHs]; 207 [DL]; 207 [DLo1]; 207 [DLo2]; 207 [DM]; 207 [DO]; 207 [Du3]; 207 [Du4]; 207 [Du5]; 207 [Du6]; 301–304, 307 [21]; 307 [27]; 307 [28] Dufresnoy, A. 450, 513 [31] Dugundji, J. 573, 585, 598 [22] Eberly, D. 170, 206 [BE] Ebmeyer, C. 319, 403 [51] Edmunds, D.E. 409, 412, 441, 513 [32] Ekeland, I. 62, 125 [39] El Hamidi, A. 341, 403 [52] Elgueta, M. 398, 403 [40] Elton, D. 491, 513 [33] Erdös, L. 478–481, 489–491, 498–502, 513 [34]; 513 [35]; 513 [36]; 513 [37]; 513 [38]; 513 [39]; 513 [40]; 513 [41] Escobar, J.F. 313, 314, 317, 344, 403 [53]; 403 [54]; 403 [55]; 404 [56]; 404 [57]; 404 [58]; 404 [59] Evans, W.D. 409, 412, 441, 491, 512 [10]; 513 [32] Fefferman, C. 453, 463, 465, 499, 500, 502, 512 [21]; 513 [22]; 513 [42]; 513 [43] Felli, V. 314, 344, 346, 404 [60]; 404 [61] Felmer, P. 385–387, 398, 403 [40]; 403 [45]; 404 [62]; 404 [63] Fernández Bonder, J. 313–315, 317, 318, 322, 328–330, 347, 348, 350, 358, 360, 361, 365, 369, 370, 376, 379, 384, 390, 391, 404 [64]; 404 [65]; 404 [66]; 404 [67]; 404 [68]; 404 [69]; 404 [70]; 404 [71]; 404 [72]; 404 [73]; 404 [74]; 404 [75]; 404 [76] Ferone, V. 61, 124 [13] Ferreira, R. 315, 376, 404 [64] Fila, M. 314, 315, 341, 343, 368, 403 [30]; 403 [31]; 403 [33]; 403 [34]; 403 [35]; 404 [77] Filo, J. 314, 404 [77] Floer, A. 526, 598 [23] Flores, C. 315, 347, 356, 360, 361, 403 [46] Flores, F. 74, 81, 124 [11] Fortunato, D. 522, 523, 594, 598 [12]; 599 [24] Fraile, J.M. 144, 207 [FKLM]; 298, 299, 307 [29] Francfort, G. 62, 123 [3] Friedrichs, K. 445, 449, 513 [44] Friesecke, G. 62, 74, 81, 97, 125 [40] Froese, R.G. 409, 425, 492, 513 [29] Fröhlich, J. 499, 500, 512 [21] Fry, M. 486, 513 [45]; 513 [46]; 513 [47] Furter, J.E. 307 [30]

603

Fusco, N. 62, 73, 123 [1]; 125 [41] Galaktionov, V. 397, 404 [78] Gámez, J.L. 298, 306 [4]; 396, 399, 400, 402 [13] Gangbo, W. 69, 124 [12] García-Azorero, J. 52 [4]; 314, 319, 322, 330, 339, 340, 350, 404 [79]; 404 [80]; 404 [81] García-Melián, J. 144, 207 [GGLS]; 300–303, 307 [31]; 307 [32] Geyler, V. 490, 513 [48] Ghoussoub, N. 3, 35, 53 [28]; 53 [52] Giachetti, D. 62, 125 [42] Gidas, B. 10, 18, 21, 53 [53]; 54 [54]; 170, 172, 179, 183, 207 [GNN]; 361, 379, 382, 404 [82]; 404 [83] Gilbarg, D. 129, 140, 207 [GT]; 240, 307 [33]; 321, 380, 384, 388, 405 [84] Glaser, V. 469, 514 [49]; 525, 598 [17] Gómez-Reñasco, R. 144, 207 [GGLS]; 300–302, 307 [31]; 307 [34]; 307 [35]; 307 [36] Graf, G.M. 499, 500, 502, 512 [21]; 513 [22] Grigor’yan, A. 466, 514 [50] Grishanov, E. 490, 513 [48] Gromov, M. 123, 125 [43] Grosse, H. 469, 514 [49] Guo, Z.M. 144, 207 [DG] Hadamard, J. 301, 307 [37] Han, Z.C. 344, 346, 405 [85]; 405 [86] Hastings, S.P. 170, 192, 207 [HM] Healey, T.J. 193, 207 [HK] Hebey, E. 313, 403 [50] Hedberg, L. 433, 436, 441, 442, 512 [4] Helffer, B. 425, 470, 514 [51]; 514 [52] Herbst, I. 474, 482, 492, 494, 512 [9] Hess, H. 431, 474, 514 [53] Hess, P. 169, 207 [He] Hilbert, D. 301, 307 [18]; 409, 513 [27] Hirsch, M. 155, 207 [Hir] Hislop, P.D. 409, 514 [54] Hoff, D. 199, 206 [CHS] Horn, R.A. 77, 113, 118, 125 [44]; 125 [45] Hsu, S.B. 168, 205, 207 [DHs]; 207 [HSW]; 207 [Hs] Hu, B. 314, 343, 379, 384, 405 [87]; 405 [88] Huang, Q. 133, 144, 207 [DH]; 301–303, 307 [28] Hundertmark, D. 463, 469, 470, 472, 514 [55]; 514 [56]; 514 [57] Hutson, V. 169, 206 [CCH]; 207 [HLM]; 207 [HLMP]; 207 [HMP] Ibrahim, M.M.A. 170, 206 [BIS] Igbida, N. 396, 402 [9] Ishii, H. 313, 405 [89]

604

Author Index

Ivrii, V. 409, 514 [58] Iwatsuka, A. 490, 514 [59] James, R.D. 121, 124 [8] Jannelli, E. 523, 557–559, 594, 599 [24]; 599 [25] Jeanjean, L. 40, 54 [55] Johnson, C.A. 77, 113, 118, 125 [44]; 125 [45] Joseph, D.D. 171, 191, 208 [JL] Kan-on, Y. 169, 200, 208 [KY]; 208 [Ka] Kantorovich, L.V. 460, 514 [60] Kato, T. 27, 53 [30]; 146, 208 [Kato]; 409, 418, 428, 429, 431, 474, 514 [61]; 514 [62]; 514 [63]; 514 [64]; 524, 546, 555, 557, 598 [10] Kavian, O. 3, 54 [56] Kawasaki, K. 169, 208 [MK] Kawohl, B. 61, 124 [13]; 364–366, 405 [90] Keller, J.B. 218, 235, 301, 307 [38] Kielhöfer, H. 3, 54 [57]; 193, 207 [HK]; 208 [K] Kirchheim, B. 79, 84, 125 [46] Kirsch, W. 409, 425, 492, 513 [29] Kishimoto, K. 169, 200, 208 [KW] Klötzler, R. 62, 125 [47] Koch-Medina, P. 144, 207 [FKLM]; 298, 299, 307 [29] Kohn, R.V. 68, 70, 116, 117, 125 [48] Koman, P. 169, 208 [KL] Kondratiev, V.A. 302, 307 [39]; 433, 450, 451, 514 [65]; 514 [66] Korman, P. 170, 208 [KLi] Krasnosel’ski˘ı, M.A. 129, 130, 208 [Kr]; 325, 330, 405 [91]; 571, 599 [26] Kryzewski, W. 51, 54 [58] Kühn, T. 473, 513 [26] Kwong, M.R. 10, 11, 54 [59] Lami Dozo, E. 315, 362–364, 369, 404 [65]; 405 [92] Laptev, A. 470–472, 514 [56]; 514 [67]; 514 [68]; 514 [69]; 514 [70] Lazer, A. 584, 587, 599 [27] Le Dret, H. 69, 115, 125 [49] Leinfelder, H. 415, 417, 432, 514 [71]; 514 [72] Leschke, H. 479, 480, 514 [73] Letelier-Albornoz, R. 302, 303, 307 [32] Leung, A.W. 169, 208 [KL] Levin, D. 459, 514 [74] Li, P. 453, 458, 514 [75] Li, S. 315, 388, 390, 406 [125] Li, S.J. 33, 34, 54 [60]; 54 [61]; 144, 207 [DL] Li, Y. 34, 54 [63]; 170, 208 [KLi]; 314, 343, 369, 405 [93]; 405 [94]

Li, Y.Y. 12, 54 [62]; 314, 344, 346, 402 [5]; 405 [85]; 405 [86] Lieb, E.H. 7, 53 [31]; 409, 452, 454, 456, 457, 462, 468–472, 482, 486, 492, 494, 497, 499, 500, 509, 512 [7]; 514 [57]; 515 [76]; 515 [77]; 515 [78]; 515 [79]; 515 [80]; 515 [81]; 515 [82]; 515 [83]; 515 [84]; 515 [85] Lin, C.S. 172, 183, 208 [LN] Lindqvist, P. 352, 405 [95] Lions, P.-L. 21, 35, 36, 39, 40, 52 [9]; 52 [25]; 53 [51]; 54 [64]; 139, 208 [L]; 313, 330, 331, 340, 368, 394, 405 [89]; 405 [96]; 405 [97]; 405 [98]; 525–527, 533, 598 [4]; 598 [8]; 599 [28]; 599 [29] Liu, Z. 24, 27, 34, 38, 42, 52 [12]; 54 [63]; 54 [65] Liu, Z.-L. 16, 24, 27, 31, 36, 38, 39, 43, 45–47, 49–51, 54 [66]; 54 [67]; 54 [68]; 54 [69] Loewner, C. 302, 308 [40] Lopes, O. 54 [70] López-Gómez, J. 144, 169, 193, 205, 206 [DLO]; 207 [FKLM]; 207 [GGLS]; 208 [LP1]; 208 [LP2]; 208 [LS]; 208 [Lop1]; 208 [Lop2]; 213, 218, 226, 252, 257, 298–305, 306 [3]; 306 [10]; 306 [11]; 306 [12]; 307 [13]; 307 [22]; 307 [23]; 307 [24]; 307 [25]; 307 [29]; 307 [30]; 307 [31]; 307 [35]; 307 [36]; 308 [41]; 308 [42]; 308 [43]; 308 [44]; 308 [45]; 308 [46]; 308 [47]; 308 [48]; 308 [49]; 308 [50]; 308 [51]; 308 [52]; 308 [53]; 308 [54]; 308 [55]; 308 [56] Lorca, S. 40, 54 [71] Loss, M. 409, 472, 478, 480, 486, 491, 499, 500, 512 [11]; 515 [79]; 515 [80]; 515 [86]; 515 [87] Lou, Y. 169–171, 189, 207 [DLo1]; 207 [DLo2]; 207 [HLM]; 207 [HLMP]; 208 [LN1]; 208 [LN2] Lundgren, T.S. 171, 191, 208 [JL] Ma, L. 137, 207 [DM] Magalhaes, C.A. 16, 53 [46] Malchiodi, A. 314, 344, 346, 402 [5] Mancebo, F.J. 313, 405 [99] Manes, A. 591, 599 [30] Maniscalco, C. 599 [31] Marcellini, P. 61, 62, 74, 76–79, 81, 83, 86, 91, 92, 94, 97, 106, 109, 114–117, 121–123, 124 [27]; 124 [28]; 124 [29]; 124 [30]; 124 [31]; 124 [32]; 125 [41]; 125 [50]; 125 [51]; 125 [52] Marcus, M. 145, 208 [MV]; 218, 301, 303, 306 [5]; 306 [6]; 306 [7]; 308 [57]; 308 [58] Marino, A. 587, 599 [32] Marshall, A.W. 77, 125 [53] Martin, A. 412, 469, 514 [49]; 515 [88]; 525, 598 [17]

Author Index Martinez, S. 315, 318, 351, 356, 360, 362, 365, 371, 376, 399, 400, 404 [66]; 404 [67]; 405 [100]; 405 [101]; 405 [102] Mascolo, E. 62, 114, 125 [54]; 125 [55]; 125 [56]; 125 [57] Masuda, K. 474, 514 [64] Matano, H. 155, 169, 208 [MM]; 208 [Ma] Matias, J. 81, 124 [9] Mawhin, J. 308 [59]; 308 [60] Mazón, J.M. 315, 370, 371, 396, 402 [8]; 402 [9]; 402 [10]; 402 [11]; 402 [12] Maz’ya, V.G. 423, 428, 433, 435, 436, 438, 439, 441, 442, 444, 446, 447, 450, 451, 514 [65]; 515 [89]; 515 [90]; 515 [91]; 515 [92]; 515 [93]; 515 [94]; 515 [95]; 515 [96]; 515 [97] McLeod, J.B. 170, 192, 207 [HM] Melgaard, M. 464, 465, 476, 482, 483, 498, 515 [98]; 515 [99]; 515 [100] Merino, S. 144, 207 [FKLM]; 298, 299, 307 [29] Merizzi, L. 27, 53 [44] Michelettii, A.M. 591, 599 [30] Milatovic, O. 421, 428, 431, 432, 512 [20] Mimura, M. 169, 170, 208 [MK]; 208 [MM]; 208 [MS]; 208 [Mi] Miranda, C. 526, 575, 599 [33] Mironescu, P. 318, 402 [21] Mischaikow, K. 169, 207 [HLM]; 207 [HLMP]; 207 [HMP] Miyagaki, O. 315, 375, 401 [1] Mohammed, A. 494, 515 [101] Molchanov, A. 449, 516 [102] Molina-Meyer, M. 218, 226, 299, 304, 306 [12]; 307 [13]; 308 [53]; 308 [54] Moll, S. 371, 402 [10] Molle, R. 526, 599 [34] Montefusco, E. 54 [72]; 342, 405 [103] Monteiro Marques, M.D.P. 62, 125 [58] Montenegro, M. 315, 395, 403 [41] Morrey, C.B. 63, 125 [59]; 125 [60] Motron, M. 315, 370, 372, 375, 405 [104] Müller, S. 62, 74, 79, 91, 123, 125 [61]; 125 [62]; 126 [63] Müller-Pfeiffer, E. 27, 54 [73] Muratori, B. 491, 512 [1]; 512 [2] Musso, M. 526, 599 [34] Nagasaki, K. 191, 208 [NS] Nakashima, K. 193, 197, 208 [N] Nash, C. 491, 512 [1]; 512 [2] Nehari, Z. 54 [74]; 525, 599 [35] Neuberger, J.M. 31, 53 [36]; 53 [37] Ni, W.-M. 10, 18, 53 [53]; 54 [75]; 169, 170, 172, 179, 183, 207 [GNN]; 208 [LN]; 208 [LN1];

605

208 [LN2]; 208 [NN]; 361, 404 [82]; 526, 598 [21] Nikishin, V.A. 302, 307 [39] Nirenberg, L. 8, 10, 12, 18, 53 [26]; 53 [32]; 53 [53]; 170, 172, 179, 183, 207 [GNN]; 302, 308 [40]; 330, 339, 361, 367, 368, 402 [22]; 404 [82]; 522, 598 [11] Nourrigat, J. 425, 514 [51] Nussbaum, R.D. 21, 53 [51]; 179, 208 [NN] Oh, Y.G. 526, 599 [36] Oleinik, I.M. 430, 431, 516 [103] Olkin, I. 77, 125 [53] Ornelas, A. 62, 125 [41]; 125 [58]; 126 [64] Ortega, R. 193, 205, 206 [DLO] Osserman, R. 218, 235, 301, 308 [61] Oswald, L. 297, 299, 306 [8] Otelbaev, M. 428, 515 [92] Ouhabaz, E.-M. 475, 476, 483, 515 [98]; 516 [104] Ouyang, T. 144, 170, 193, 207 [DO]; 208 [Ou]; 209 [OS1]; 209 [OS2]; 298, 308 [62] Oxtoby, J.C. 84, 126 [65] Pacella, F. 17, 31, 33, 52 [1]; 54 [76]; 340, 405 [98] Palais, R.S. 54 [77] Palmieri, G. 522, 598 [12] Pankov, A. 41, 52 [13] Pao, C.V. 169, 209 [Pao] Pardo, R.M. 205, 208 [LP1]; 208 [LP2] Park, Y.J. 313, 314, 344, 405 [105]; 405 [106] Parter, S.V. 170, 209 [Pa] Passaseo, D. 17, 52 [23]; 526, 599 [34] Peletier, L.A. 395, 396, 403 [38]; 522–524, 594, 598 [2]; 598 [3] Peller, V.V. 473, 516 [105] Peral, I. 52 [4]; 314, 319, 322, 330, 339, 340, 350, 404 [79]; 404 [80]; 404 [81] Perrotta, S. 61, 124 [14]; 124 [15] Pflüger, K. 339, 405 [107] Phillips, D. 61, 124 [10] Pianigiani, G. 74, 81, 125 [38] Pierrotti, D. 315, 339, 375, 405 [108]; 405 [109] Pinasco, J.P. 315, 390, 404 [68]; 404 [69] Pisante, G. 62, 69, 79, 84, 86, 92, 111, 113, 117, 124 [33]; 125 [34]; 126 [66] Pitt, L.D. 443, 473, 516 [106] Pohozaev, S. 6, 54 [78]; 522, 541, 599 [37] Polacik, P. 169, 207 [HLMP]; 207 [HMP] Povzner, A.Ya. 426, 427, 516 [107] Prodi, G. 587, 599 [32] Protter, M.H. 129, 195, 197, 209 [PW]; 301, 308 [63]

606

Author Index

Quittner, P. 305, 308 [55]; 308 [64]; 314, 341, 342, 403 [31]; 403 [32] Rabinowitz, P.H. 3, 5, 31, 33, 36, 45–49, 52 [5]; 53 [48]; 54 [79]; 129–131, 177, 189, 193, 206 [CR1]; 206 [CR2]; 209 [Ra]; 237, 239, 307 [19]; 307 [20]; 325, 326, 329, 390, 402 [6]; 405 [110]; 526, 598 [18]; 599 [38] R˘adulescu, V.D. 302, 303, 307 [14]; 307 [15]; 307 [16]; 307 [17]; 342, 403 [36]; 403 [37]; 405 [103]; 494, 515 [101] Raikov, G.D. 494, 498, 515 [101]; 516 [108]; 516 [109]; 516 [110]; 516 [111] Ramos, M. 54 [80] Raoult, A. 69, 115, 125 [49] Raymond, J.P. 62, 126 [67]; 126 [68]; 126 [69]; 126 [70] Reed, M. 409, 414, 417–420, 422, 423, 428, 429, 446, 456, 460, 490, 516 [112]; 516 [113]; 516 [114] Reichel, W. 315, 368, 403 [35]; 405 [111] Ribeiro, A.M. 62, 69, 76, 82, 86, 92, 111, 113, 114, 117, 125 [34]; 125 [35] Robert, D. 470, 514 [52] Robinson, S.B. 358, 399, 403 [49] Rockafellar, R.T. 82, 96, 126 [71] Rofe-Beketov, F.S. 430, 516 [115] Rossi, J.D. 313–315, 317–319, 321, 322, 328–330, 339, 344, 347, 348, 350, 351, 356, 358, 360–362, 365, 369–371, 376, 379, 384, 390, 391, 395, 397–400, 402 [11]; 402 [14]; 403 [42]; 404 [64]; 404 [65]; 404 [66]; 404 [67]; 404 [68]; 404 [69]; 404 [70]; 404 [71]; 404 [72]; 404 [73]; 404 [74]; 404 [75]; 404 [76]; 404 [81]; 405 [100]; 405 [101]; 405 [102] Rothe, F. 205, 206 [dMR] Rozenblum, G. 409, 410, 412, 428, 452, 462–465, 476, 482–485, 498, 515 [98]; 515 [99]; 515 [100]; 516 [116]; 516 [117]; 516 [118]; 516 [119]; 516 [120]; 516 [121] Ruder, R. 479, 480, 514 [73] Ruijsenaars, S. 421, 516 [122] Sabina de Lis, J.C. 144, 169, 207 [GGLS]; 208 [LS]; 300–303, 307 [31]; 307 [32]; 308 [56] Sacks, P. 315, 396, 402 [19] Safarov, Yu. 409, 516 [123] Sakamoto, K. 170, 208 [MS] Sattinger, D.H. 132, 209 [Sa]; 232, 308 [65] Schechter, M. 3, 16, 54 [81]; 55 [82]; 315, 388, 390, 406 [112] Schianchi, R. 62, 114, 125 [42]; 125 [55]; 125 [56]; 125 [57]

Schrader, R. 431, 474, 514 [53] Schwinger, J. 446, 516 [124] Segura de Leon, S. 315, 396, 402 [12] Shafrir, I. 314, 343, 403 [30]; 403 [33] Shaposhnikova, T.O. 433, 436, 441, 442, 515 [93] Shen, Z. 499, 500, 516 [125]; 516 [126] Shi, J. 170, 193, 209 [OS1]; 209 [OS2]; 209 [S] Shigekawa, I. 490, 516 [127] Shivaji, R. 170, 206 [BIS]; 206 [CS]; 209 [Sh] Shubin, M. 409, 410, 412, 420, 421, 425–428, 430–433, 446, 450, 451, 512 [13]; 512 [20]; 514 [65]; 514 [66]; 515 [94]; 516 [120]; 516 [128]; 516 [129] Sigal, I.M. 409, 514 [54] Simader, C.G. 415, 417, 432, 514 [72] Simon, B. 299, 309 [66]; 409, 414, 417–425, 428, 429, 446, 452, 456, 460, 462, 467, 473, 474, 478, 482, 490, 492, 494, 512 [9]; 513 [29]; 516 [112]; 516 [113]; 516 [114]; 517 [130]; 517 [131]; 517 [132]; 517 [133]; 517 [134]; 517 [135]; 517 [136]; 517 [137]; 517 [138] Simon, J. 323, 406 [113] Simondon, F. 305, 308 [64] Smets, D. 17, 55 [83] Smith, H.L. 155, 207 [HSW]; 209 [Sm] Smoller, J. 199, 206 [CHS] Sobolev, A.V. 494, 498–500, 517 [139]; 517 [140]; 517 [141]; 517 [142]; 517 [143]; 517 [144] Sola-Morales, J. 314, 344, 402 [24]; 403 [39] Solimini, S. 521, 523–529, 532, 557–559, 570, 571, 581, 584, 587, 594, 598 [13]; 598 [14]; 598 [19]; 598 [20]; 599 [25]; 599 [27]; 599 [39]; 599 [40]; 599 [41]; 599 [42]; 599 [43]; 599 [44] Solomyak, M. 409, 410, 412, 414, 456, 459, 462–464, 482, 507, 511, 512 [16]; 512 [17]; 512 [18]; 514 [74]; 516 [120]; 516 [121] Solovej, J.P. 486, 491, 492, 494, 497, 499–502, 513 [36]; 513 [37]; 513 [38]; 513 [39]; 513 [40]; 515 [80]; 515 [81]; 515 [82]; 515 [83] Solynin, P. 9, 17, 53 [34] Spruck, J. 21, 54 [54]; 379, 382, 404 [83] Srikanth, P.N. 10, 55 [84] Stampacchia, G. 319, 321, 406 [114] Steklov, M.W. 313, 317, 329, 350, 406 [115] Strang, G. 68, 70, 116, 117, 125 [48] Strauss, W.A. 9, 55 [85] Struwe, M. 3, 33, 35, 55 [86]; 55 [87]; 55 [88]; 523, 527, 530, 539, 571, 581, 584, 594, 598 [14]; 599 [45]; 599 [46] Stubbe, J. 472, 499, 500, 512 [19]; 512 [21] Suárez, A. 303, 304, 307 [23]; 307 [24]; 307 [25] Sun, J. 24, 54 [65]

Author Index Suzuki, T. 191, 208 [NS] Šverák, V. 62, 63, 74, 123, 125 [61]; 125 [62]; 126 [72]; 126 [73] Sweers, G. 396, 406 [116] Sychev, M.A. 62, 79, 91, 126 [63]; 126 [74]; 126 [75] Szulkin, A. 51, 54 [58] Tachizawa, K. 465, 469, 517 [145] Tahraoui, R. 61, 62, 123 [4]; 123 [5]; 124 [6]; 126 [76]; 126 [77] Taira, K. 209 [T] Tamura, H. 421, 498, 517 [146]; 517 [147]; 517 [148] Tanaka, K. 40, 54 [55] Tang, M. 193, 209 [Tang] Tanteri, C. 62, 76, 83, 113, 117, 123, 125 [36]; 125 [37] Tehrani, H. 35, 53 [28] Terracini, S. 27, 31, 53 [44]; 53 [45]; 314, 315, 339, 343, 375, 405 [108]; 405 [109]; 406 [117] Teta, R. 421, 512 [3] Thaller, B. 425, 478, 480, 492, 515 [86]; 517 [149] Thangavelu, S. 493, 517 [150] Thayer, J. 389, 406 [118] Thirring, W. 454, 468–472, 509, 515 [84]; 515 [85] Thomas, L.E. 469, 470, 472, 514 [57] Toledo, J. 315, 396, 402 [9]; 402 [12] Tolksdorf, P. 321, 347, 356, 376, 406 [119] Torne, O. 315, 362–364, 405 [92] Treu, G. 62, 126 [78] Tricarico, M. 340, 405 [98] Troestler, C. 51, 55 [89] Trudinger, N.S. 129, 140, 207 [GT]; 240, 307 [33]; 321, 380, 384, 388, 405 [84]; 406 [120] Turner, R.E.L. 21, 53 [33] Ubilla, P. 40, 54 [71] Uhlenbrock, D.A. 431, 474, 514 [53] Umezu, K. 314, 342, 406 [121]; 406 [122] Ural’ceva, N. 428, 517 [151] van Heerden, F.A. 27, 43, 54 [66] Vassiliev, D. 409, 516 [123] Vázquez, J.L. 347, 352, 354, 376, 397, 398, 404 [78]; 406 [123] Vega, J.M. 313, 405 [99] Verbitsky, I.E. 433, 436, 438, 439, 442, 444, 515 [95]; 515 [96]; 515 [97]; 517 [152]

607

Véron, L. 145, 208 [MV]; 218, 301, 303, 308 [57]; 308 [58]; 309 [67] Verzini, G. 31, 53 [45] Vugalter, V. 489–491, 513 [41] Waltman, P. 207 [HSW] Wang, S.-H. 170, 209 [W1]; 209 [W2] Wang, X. 425, 514 [51] Wang, Z.-Q. 12, 16, 24, 27, 31, 33, 34, 36, 38, 39, 41, 43, 45–47, 49–51, 52 [13]; 52 [14]; 52 [15]; 52 [16]; 53 [38]; 53 [47]; 54 [60]; 54 [61]; 54 [66]; 54 [67]; 54 [68]; 54 [69]; 54 [80]; 55 [90]; 55 [91]; 55 [92]; 55 [93]; 387, 404 [63]; 526, 598 [5] Warzel, S. 479, 480, 498, 514 [73]; 516 [111] Wei, J. 33, 55 [94] Wei, J.C. 18, 54 [75] Weidl, T. 456, 469, 471, 472, 514 [56]; 514 [69]; 514 [70]; 517 [153]; 517 [154] Weidmann, J. 414, 420, 490, 517 [155] Weinberger, H.F. 129, 169, 195, 197, 200, 208 [KW]; 209 [PW]; 301, 308 [63] Weinstein, A. 526, 598 [23] Weth, T. 17, 24, 27, 31, 33, 34, 38, 42, 52 [12]; 52 [17]; 52 [18]; 52 [19]; 523, 594, 598 [15] Wheeden, R.L. 465, 513 [23] Wiebers, H. 170, 209 [Wie1]; 209 [Wie2] Willem, M. 3, 8, 9, 12, 15, 17, 31, 40, 51, 52 [19]; 52 [20]; 54 [80]; 55 [83]; 55 [89]; 55 [93]; 55 [95]; 55 [96]; 525, 598 [6] Winter, M. 33, 55 [94] Wolanski, N. 315, 365, 404 [76] Yadava, S.L. 344, 361, 368, 369, 402 [3] Yamabe, H. 298, 309 [68] Yamada, Y. 205, 209 [Y] Yanagida, E. 169, 208 [KY] Yang, J. 336, 402 [27] Yau, H.-T. 486, 491, 499, 515 [87] Yau, S.-T. 453, 458, 466, 514 [50]; 514 [75] Yin, H.M. 379, 405 [88] Yngvason, J. 492, 494, 497, 515 [81]; 515 [82]; 515 [83] Yosida, K. 84, 126 [79] Zagatti, S. 62, 114, 124 [19]; 126 [80] Zhu, M. 314, 343, 346, 369, 405 [93]; 405 [94]; 406 [124] Zou, W. 16, 55 [82]; 315, 388, 390, 405 [111]; 406 [112]; 406 [125]

This page intentionally left blank

Subject Index

CLR estimate, 412 compact operator, 503 competition system, 193 concentration–compactness, 521, 525 – method, 330, 340 constant electric field, 429 continuous spectrum, 504 convergence quasieverywhere, 434 convex – envelope, 73 – – , definition, 67 – hull, 74, 81 – – , definition, 75 – integration, 123 Coulomb gauge, 415 coupling constant, 412 Crandall and Rabinowitz transversality condition, 237 critical – dimension, 368 – exponent, 315, 322, 340, 343, 367 critical growth, 521, 522, 524, 526, 527, 529, 539, 548, 587 crowding effects, 216 Cwikel – inequality, 455 – – , operator-valued case, 463

A a priori bounds of Keller and Osserman, 218, 235 adjoint operator, 504 Aharonov–Bohm – effect, 415 – magnetic potential, 421 Aharonov–Casher theorem, 487 Alibert–Dacorogna–Marcellini example, 65 anticommutation relation, 416 antimaximum principle, 395 approximation property, 81, 115 – , definition, 80 attracting property, 223 B Baire category theorem, 79, 84 band subspace, 493 Bessel capacity, 434 bifurcation – from a simple eigenvalue, 130 – from infinity, 131 Birman–Schwinger – operator, 440, 441, 453, 511 – principle, 446, 510 blow-up technique, 382 Bohr–Sommerfeld quantization condition, 411 Bolza example, 60, 106 boundary blow-up, 133

D Dacorogna formula, 68 decay estimates, 541 deformation lemma, 14 diamagnetic – inequality, 474 – – , strong, 443 – – , weak, 424 – monotonicity, 477 diamagneticity, 424 differential – expression, 414 – operation, 414 distribution function, 506

C Carathéodory theorem, 68, 102, 103, 106 classical – constant, 411 – Hamiltonian, 409 – logistic equation, 234, 235 – solution, 219, 223 closable – form, 420 – – , relatively, 422 – operator, 503 closed operator, 503 closure, 417 609

610 domain of the operator, 503 domination, 443 double natural constraint, 570, 581, 594, 596 E eigenvalue cluster, 498 eigenvalues, 411 electric potential, 416 elliptic system, 378 embedding theorem, 419 equivalent constants, 437 essential spectrum, 505 essentially self-adjoint operator, 417, 504 evolution equation, 426 exchange stability principle, 239 explosive solution, 216, 218 F Feynman–Kac formula, 456 finite propagation speed property, 427 Fredholm analytic pencil, 236 free boundary, 395 Friedrichs extension, 421 Fuˇcík spectrum, 400, 401 function of first class, 84 G gauge transformation, 414 Gelfand equation, perturbed, 169 genus, 325 geodesically complete manifold, 427 Glazman lemma, 494, 508 gradient elliptic systems, 390 group action, 11 growth rate – , boundary, 278, 279 – , intrinsic, 216 H Hamiltonian elliptic systems, 385 harmonic oscillator, 419 Harnack inequality, 300 Hausdorff measure, 434 heterogeneous competition system, 147 Holmgren principle, 427 I indefinite superlinear parabolic problems, 305 infinitesimal form-boundedness, 433 J Jensen inequality, 73, 81, 103

Subject Index K K-attractive, 26 Kato inequality, 431 Kato–Rellich theorem, 418 KLMN theorem, 422 L lamination convex hull, definition, 75 Landau – bands, 493 – levels, 487, 492 Landesman–Lazer condition, 399, 400 Laplace–Beltrami operator, 414 Laplacian, 414 – , magnetic, 414 large – solution, 216, 218, 220, 278 – – , continuity, 296 – – , strong monotonicity, 295 layer solution, 344 Lieb–Thirring inequality, 413 – for operator-valued potentials, 471 – for Pauli operators, 492, 498 Lieb–Thirring inequality, 468, 494 linear operator, 503 localization method, 303 M magnetic – bottle, 450 – field, 414 – gradient, 414 – potential, 414 Malthus law, 216, 217 maximal – operator, 418 – solution, 219, 226 metasolution, 216, 220, 223 method of moving spheres, 343 min–max, 521, 523, 570, 578, 580, 581, 583, 584, 588–590, 597 minimal – operator, 419 – solution, 219, 226 Molchanov functional, 449 mountain pass theorem, 333 multibump solution, 45 multiplicity, 521, 523, 524, 526, 527, 536, 539, 541, 570, 593, 594 – , exact, 169 N Nehari manifold, 5, 13 Neumann–Schatten class, 508

Subject Index nodal – domain, 27 – Nehari set, 22 – properties, 193 nod(u), 27 nonlinear boundary conditions, 313, 314, 317, 319, 322, 323, 378 nonvariational systems, 378 null Lagrangian, 63 P p-capacity, 367 Palais–Smale condition, 323, 324, 328, 330, 331, 334, 335, 394 paramagnetic monotonicity, 477 patterned solution, 137 Pauli – matrices, 415 – operator, 413, 415 phase-space volume, 452, 467 Pohozaev – formula, 546, 550 – identity, 338, 361, 522, 523, 541, 544, 546 – inequality, 527, 541, 544 Poincaré gauge, 479 point spectrum, 504 polyconvex – envelope, definition, 67 – function, 64, 66, 95, 99, 100 – – , definition, 62 – hull, definition, 75 population behavior, 217 porous medium equation, 304 positively – dominated, 473 – invariant, 24 – – , strictly, 24 – preserving, 443, 460 – – semigroups, 460 potential wells, 121, 122 predator–prey system, 200 Q quadratic form, 419 – , closed, 419 quasiaffine function, 63–65, 68, 87–89, 110, 112–114, 116–118, 120 – , definition, 62 quasiconvex – envelope, 69, 71, 86, 115, 117 – – , definition, 67 – function, 64, 72, 78, 79, 86, 92 – – , definition, 62 – hull, definition, 75

611

R Rabinowitz – theorem, 343 – – , global bifurcation, 130 rank one convex – envelope, definition, 67 – function, definition, 62 – hull, 76 – – , definition, 75 relatively – bounded operator, 418 – – , infinitesimally, 418 – compact forms, 436 relaxation – property, 77, 78, 80, 81, 87–91, 112, 113, 115, 117 – – , definition, 77 – theorem, 60, 70, 86 resolvent – operator, 236 – set, 504 resonance, 399, 400 Riesz capacity, 434 S s-numbers, 507 Saint Venant–Kirchhoff energy, 69 Schatten ideals, 473 Schrödinger – equation, 411 – operator, 409 – – , magnetic, 431, 450 Schwarz – symmetric, foliated, 17, 31 – symmetrization, 9 – – , foliated, 17 self-adjoint – extension, 417 – operator, 504 separability, 428 separation theorem, 64, 68 sesquilinear form, 420 simple eigenvalue, 131 singular values, 69, 76, 77, 82, 110, 111, 115 – , definition, 65 Sobolev – constant, 357 – embedding, 320, 358, 361, 534, 537, 539, 565 – inequality, 423 – space, 316, 317, 417 – trace, 317, 340 – – constant, 316, 317, 355, 356, 365, 367, 371, 375, 377

612 – – embedding, 318, 331, 358, 360, 367 – – inequality, 322, 351 – – theorem, 313, 316, 326, 333, 361 spatial degeneracy, 131 spectral – measure, 505 – theorem, 505 spectrum, 504 spherical symmetrization, 364, 366 spherically symmetric, 364, 366 stability – analysis, 154 – of matter, 468 stable pattern, 163 Stark potential, 419 Steklov eigenvalues, 313, 317, 329, 350 strictly convex function in at least N directions, 95, 97, 101 – , definition, 94 strictly quasiconvex function, 94, 97, 99 – , definition, 92 strong maximum principle, 226 strongly increasing, 256 structural stability, 240 Stummel class, 428 subcritical – exponent, 322, 342 – growth, 521, 524 subsolution, 226

Subject Index supersolution, 226 – element, 254 – positive, strict, 226 Šverák example, 63 symmetric operator, 504 symmetry and symmetry breaking, 362 T trace class operator, 507 trace-property, 371 Trotter–Kato–Masuda formula, 474 turning point theorem, 131 U unbounded domains, 533 V Verhulst law, 216, 217 W weak lower semicontinuity, definition, 64 Wiener – capacity, 433 – integral, 456 Y Yamabe problem, 344 Z zero modes, 413, 486 – density, 491